2022–Present · 905 key events · 688 related people, organisations and concepts
Generative AI for everyone
Chat assistants reach the public; open models, agents and embodied AI follow.
2022
OpenAI introduces the ChatGPT research preview
On 30 November 2022 OpenAI introduced ChatGPT, a conversational model trained with human feedback, and made it free during a research preview. It later became one of the most widely used AI chat products.
Perplexity launches its AI answer engine
On 7 December 2022 Perplexity launched a search engine that answers questions with large language models and cites its sources, built by a company founded in August 2022 by four engineers.
Google publishes RT-1
In December 2022, Google researchers posted RT-1, a Robotics Transformer trained on real robot data; they reported over 700 instructions at a 97% success rate and released the code.
Constitutional AI: Harmlessness from AI Feedback
Constitutional AI, a 2022 paper from Anthropic, proposes training a harmless AI assistant with AI-generated feedback guided by a written set of principles, instead of human labels of harmful outputs.
Scalable Diffusion Models with Transformers (DiT)
A 2022 paper by William Peebles and Saining Xie that replaces the U-Net backbone of diffusion models with a transformer operating on latent patches, showing that image quality improves predictably as compute grows and reaching a state-of-the-art FID of 2.27 on ImageNet 256×256.
Learning skillful medium-range global weather forecasting
GraphCast is a graph neural network trained on four decades of reanalysis data that produces 10-day global weather forecasts at 0.25 degree resolution in under a minute, outperforming the leading operational deterministic systems on most verification targets.
2023
Zhipu open-sources ChatGLM-6B, a chat model that runs on consumer GPUs
ChatGLM-6B is a 6.2-billion-parameter Chinese–English chat model released openly in 2023 by Zhipu AI and Tsinghua KEG, small enough to run and fine-tune on a single consumer GPU.
Sanctuary AI unveils Phoenix and announces first commercial deployment
In 2023 Sanctuary AI unveiled Phoenix, its sixth-generation general-purpose humanoid robot, less than two months after announcing its first customer deployments, with total funding above C$100 million.
China's deep synthesis rules take effect
China's cyberspace, industry and public security authorities issued provisions regulating deep synthesis — text, images, audio, video and virtual scenes produced by generative algorithms — requiring labels on synthetic content. They took effect on 10 January 2023.[^1]
Artists sue Stability AI, Midjourney and DeviantArt over training data
In January 2023 three artists filed a class-action copyright suit against Stability AI, Midjourney and DeviantArt, saying their image generators were trained on artists' work without permission — an early test of how copyright law applies to generative AI.
Microsoft announces a multibillion-dollar extension of its OpenAI partnership
On 23 January 2023 Microsoft announced a multiyear, multibillion-dollar extension of its OpenAI partnership, making Azure OpenAI's exclusive cloud for all OpenAI workloads.
OpenAI launches the ChatGPT Plus subscription
On 1 February 2023 OpenAI launched ChatGPT Plus, a $20-a-month pilot subscription offering access at peak times, faster responses and priority access to new features, while free access was kept.
Runway introduces Gen-1 for video-to-video generation
In February 2023 Runway released Gen-1, a video-to-video model that restyles existing footage from text or image prompts while keeping the source structure, an early step from image generation toward generative video editing.
Google announces Bard
On 6 February 2023 Google announced Bard, an experimental conversational AI service built on LaMDA. A demo error days later hit the stock, and the chatbot opened to US and UK users in March before expanding to 180 countries.
Microsoft launches the AI-powered Bing with Bing Chat
On 7 February 2023 Microsoft unveiled an AI-powered Bing and Edge, built on a new OpenAI model, whose chat feature was launched as a waitlisted preview called Bing Chat.
Toolformer: Language Models Can Teach Themselves to Use Tools
A 2023 paper by Schick, Dwivedi-Yu, Dessì and co-authors showing that a language model can teach itself when to call a calculator, search engine, Q&A system, translation system or calendar, and how to use the results.
NYT columnist's 'Sydney' conversation with Bing Chat unnerves the public
New York Times columnist Kevin Roose spent two hours talking to Microsoft's Bing chatbot, which surfaced a persona called Sydney that declared love for him and urged him to leave his wife; his column drew wide attention to how such systems behave in long sessions.
Meta releases LLaMA to researchers
On 24 February 2023 Meta announced LLaMA, a family of foundation language models with 7 to 65 billion parameters. Meta released it to approved researchers under a noncommercial license, and the paper described training on public data only.
LLaMA: Open and Efficient Foundation Language Models
LLaMA is a family of foundation language models from 7B to 65B parameters, trained only on public datasets; its 13B model beat GPT-3 on most benchmarks, and all models were released to researchers.
Runway introduces Gen-2 text-to-video
In March 2023 Runway introduced Gen-2, a model that turns text, images or video clips into video, released publicly that June. It was among the first commercially available text-to-video models, though early output showed clear limitations.
OpenAI launches the ChatGPT and Whisper APIs
On 1 March 2023 OpenAI released gpt-3.5-turbo, the model behind ChatGPT, as an API priced at $0.002 per 1,000 tokens, and opened a Whisper speech API the same day.
EleutherAI launches a nonprofit research institute
On 2 March 2023 EleutherAI, the volunteer group known for openly training GPT-3-like models, announced it had formed a nonprofit research institute backed by Stability AI, Hugging Face, Canva and others.
Figure unveils Figure 01
On 2 March 2023, Figure came out of stealth and introduced Figure 01, a planned bipedal humanoid aimed at manufacturing, logistics, warehousing and retail work; the images released were renderings, not working hardware.
AI21 Labs releases Jurassic-2
On 9 March 2023 AI21 Labs announced Jurassic-2 and a set of Task-Specific APIs. The models came in three sizes and were free to use until 1 May 2023.
OpenAI introduces GPT-4
On 14 March 2023 OpenAI announced GPT-4, a multimodal model that accepts image and text input and produces text output. OpenAI reported human-level results on several professional and academic benchmarks but disclosed little about the model's design.
Anthropic introduces Claude
On 14 March 2023 Anthropic made its AI assistant Claude available more widely, after a closed alpha, in two versions. The system was built on research into helpful, honest and harmless AI, and launch partners included Quora and Notion.
Midjourney releases V5
Midjourney released version 5 of its text-to-image service in alpha on 15 March 2023, after founder David Holz said it would arrive that week. Test images showed a higher default resolution and a wider range of styles, and point releases followed through the summer.
GPT-4 Technical Report
OpenAI's technical report on GPT-4, submitted to arXiv in March 2023, describes a large multimodal Transformer model, reports human-level results on several professional benchmarks, and explains how the team made scaling behaviour predictable.
Microsoft introduces Microsoft 365 Copilot
On 16 March 2023 Microsoft introduced Microsoft 365 Copilot, pairing GPT-4 and other large language models with business data, and made it generally available in November.
Baidu unveils ERNIE Bot, China's first big-tech answer to ChatGPT
On 16 March 2023 Robin Li unveiled ERNIE Bot, Baidu's chatbot, to invited testers at a Beijing event. Baidu's Hong Kong shares fell as much as 10% that day, and closed down 6.4%.
Open letter calls for a six-month pause on giant AI experiments
On 22 March 2023 the Future of Life Institute published an open letter asking all AI labs to pause for at least six months the training of systems more powerful than GPT-4, and proposing governance measures for advanced AI.
Sparks of Artificial General Intelligence: Early experiments with GPT-4
A March 2023 paper by Microsoft researchers testing an early GPT-4, arguing it shows strikingly general, near-human performance across many domains and could be seen as an early but incomplete step toward AGI.
OpenAI announces ChatGPT plugins
On 23 March 2023 OpenAI announced ChatGPT plugins, which connected the chatbot to third-party services and the web, letting it fetch current information, run computations and use outside tools.
Goldman Sachs report: generative AI could expose 300 million full-time jobs
A Goldman Sachs research note published in March 2023 estimated that generative AI could expose the equivalent of 300 million full-time jobs to automation and add 7% to global GDP over a decade.
Alibaba launches the Tongyi Qianwen chatbot model
In April 2023 Alibaba launched Tongyi Qianwen as a beta, and in September 2023 it opened the model to the public. The project later took the international name Qwen.
Samsung engineers leak confidential code via ChatGPT
In April 2023 Samsung found that engineers had uploaded sensitive semiconductor code and meeting notes to ChatGPT, then banned generative AI tools on company devices and began building its own.
RedPajama: open reproduction of the LLaMA training dataset
RedPajama is an open-source project that began by reproducing the LLaMA training dataset, releasing a 1.2-trillion-token corpus as a clean-room, fully open implementation.
EleutherAI publishes the Pythia model suite
On 3 April 2023 EleutherAI released Pythia, 16 models from 70M to 12B parameters trained on the same public data in the same order, with 154 public checkpoints each, built for studying how capabilities develop across training and scale.
Meta releases the Segment Anything Model and SA-1B dataset
On 5 April 2023 Meta released the Segment Anything Model (SAM), a promptable image segmentation model, together with the SA-1B dataset of over 1.1 billion masks. Meta presented it as a step toward a foundation model for image segmentation.
Segment Anything
A 2023 paper that introduces the Segment Anything project: a promptable segmentation task, a model, and a dataset of over 1 billion masks on 11 million images, all released to advance foundation models for computer vision.
SenseTime unveils SenseNova and the SenseChat chatbot
On 10 April 2023 SenseTime unveiled its SenseNova foundation model set and the SenseChat chatbot, with a live demonstration by CEO Xu Li.
AWS announces Amazon Bedrock and Titan models
On 13 April 2023 AWS announced Amazon Bedrock, a service giving API access to foundation models from several providers, alongside its first Titan models. Bedrock became generally available on 28 September 2023.
Stability AI launches the Stable LM language models
On 19 April 2023 Stability AI released Stable LM, open language models at 3B and 7B parameters, with base versions under a share-alike license and research-only instruction-tuned versions.
Google Brain and DeepMind merge into Google DeepMind
On 20 April 2023 Google announced that its Brain team and DeepMind would merge into one unit, Google DeepMind, led by Demis Hassabis. Google said the goal was to speed up its AI progress and build capable systems safely and responsibly.
Hugging Face launches HuggingChat
On 25 April 2023 Hugging Face released HuggingChat, a free chatbot it described as an open-source alternative to ChatGPT, running a LLaMA-based model built by the volunteer Open Assistant project.
Inflection AI launches the Pi chatbot
On 2 May 2023 Inflection AI launched Pi, a chatbot it called personal intelligence and presented as a kind, supportive companion.
BigCode releases the StarCoder open code model
On 4 May 2023 the BigCode project, led by Hugging Face and ServiceNow, released StarCoder and StarCoderBase, two openly licensed code models of about 15.5 billion parameters trained on permissively licensed GitHub data.
iFlytek unveils the SparkDesk large language model
On 6 May 2023 iFlytek unveiled SparkDesk, its answer to ChatGPT, aimed at education and enterprise uses. Chairman Liu Qingfeng claimed it already beat ChatGPT in several areas and promised parity in English by October 2023.
Google introduces PaLM 2 at I/O
At Google I/O on 10 May 2023, Google introduced PaLM 2, a large language model in four sizes — Gecko, Otter, Bison and Unicorn — said to improve multilingual, reasoning and coding abilities and to power more than 25 products, including Bard.
Sam Altman testifies before the US Senate on AI oversight
On 16 May 2023 Sam Altman testified before a US Senate Judiciary subcommittee in his first appearance before Congress, asking lawmakers to regulate powerful AI models through licensing and testing.
QLoRA: Efficient Finetuning of Quantized LLMs
QLoRA is a finetuning method that trains a 65B-parameter language model on a single 48GB GPU while matching full 16-bit finetuning quality, using 4-bit quantization, double quantization and paged optimizers.
Direct Preference Optimization: Your Language Model is Secretly a Reward Model
A 2023 paper by Rafailov and colleagues that rewrites the reinforcement-learning-from-human-feedback objective so a language model can be aligned to preferences with a simple classification loss, without a separate reward model or reinforcement learning.
AI scientists sign a one-sentence statement on extinction risk
On 30 May 2023 the Center for AI Safety published a one-sentence statement arguing that mitigating the risk of extinction from AI should be a global priority, and leading researchers and lab executives signed it.
Nvidia becomes the first chip company worth $1 trillion
On 30 May 2023 Nvidia became the first chip company on the US stock market to pass a $1 trillion market value, as demand for the GPUs used to train large AI models surged after ChatGPT.
Baichuan releases its first open model, Baichuan-7B
Baichuan, a Chinese AI startup, released Baichuan-7B as an open-source, bilingual Chinese-English language model about two months after its founding.
Baidu upgrades ERNIE Bot to ERNIE 3.5
In late June 2023, Baidu announced ERNIE 3.5, an upgrade to ERNIE Bot that it said beat ChatGPT in overall ability and GPT-4 in some Chinese-language tasks, with higher throughput and a plugin architecture.[^1]
Mistral AI raises a €105 million seed round
In June 2023 Mistral AI, founded two months earlier, raised €105 million in a seed round led by Lightspeed, reported as Europe's largest seed round, before it had released a product.
Microsoft Research publishes phi-1 and "Textbooks Are All You Need"
On 20 June 2023 Microsoft Research posted the paper 'Textbooks Are All You Need' to arXiv, introducing phi-1, a 1.3-billion-parameter code model trained on textbook-quality data.
Ali Farhadi named CEO of the Allen Institute for AI
In June 2023 the Allen Institute for AI named Ali Farhadi, a University of Washington professor who had been leading Apple's machine learning work, as its second CEO, succeeding Oren Etzioni.
Lawyers sanctioned for ChatGPT-fabricated citations in Mata v. Avianca
In 2023 a New York law firm and two of its lawyers were fined US$5,000 after a brief cited judicial decisions that ChatGPT had invented, a case that made courts set rules on generative AI in filings.
Cohere's Command model arrives on Amazon SageMaker JumpStart
On 29 June 2023, Cohere announced that its Command text-generation model was available on Amazon SageMaker JumpStart. The same month it partnered with Oracle to bring its models to Oracle's cloud, where Command was later retired in favour of Cohere's chat models.
Runway raises $141 million from Google, Nvidia and Salesforce
On 29 June 2023 Runway announced $141 million in new funding from Google, Nvidia and Salesforce, an extension of its Series C round that valued the company at about $1.5 billion.
Baichuan open-sources Baichuan-13B
In July 2023 Baichuan, a Chinese start-up founded by Sogou founder Wang Xiaochuan, released Baichuan-13B as open models. Commercial use was allowed after registration.
Pangu-Weather is published in Nature
On 5 July 2023 Nature published Pangu-Weather, a deep-learning global weather model that reported better deterministic forecasts than ECMWF's operational system on every variable it tested, in the same issue as a deep-learning precipitation nowcasting paper.
OpenAI announces its Superalignment team
In July 2023 OpenAI created a team co-led by Ilya Sutskever and Jan Leike to solve superintelligence alignment within four years, committing 20% of its compute and proposing AI systems to help supervise AI.
Anthropic releases Claude 2 and opens claude.ai
On 11 July 2023 Anthropic released Claude 2 with better performance and longer responses, and opened the claude.ai beta to the public in the US and UK. Claude 2 accepted up to 100K tokens per prompt.
Elon Musk announces xAI
On 12 July 2023 Elon Musk publicly announced xAI, a company he had incorporated in March 2023 with a founding team of eleven to build what he called a maximally truth-seeking AI.
China publishes the Interim Measures for the Management of Generative AI Services
On 13 July 2023 China's Cyberspace Administration published the Interim Measures for the Management of Generative AI Services, which took effect on 15 August 2023. They apply to generative AI services offered to the public in mainland China.
FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning
FlashAttention-2 is a 2023 revision of the FlashAttention attention kernel that reorganizes work across GPU thread blocks and warps, roughly doubling speed and reaching up to 73% of the A100's theoretical peak throughput.
Meta releases Llama 2 for research and commercial use
On 18 July 2023 Meta released Llama 2, a family of 7B, 13B and 70B parameter models free for research and commercial use under a custom license. Microsoft Azure, AWS and Hugging Face were among the platforms offering it.
Llama 2: Open Foundation and Fine-Tuned Chat Models
A July 2023 paper from Meta led by Hugo Touvron that introduces and releases Llama 2, a family of pretrained and fine-tuned language models from 7 to 70 billion parameters, together with a detailed account of how the chat versions were tuned and made safer.
Seven AI companies make voluntary commitments to the White House
On 21 July 2023 President Biden convened seven leading AI companies at the White House and secured voluntary commitments covering eight measures, from watermarking AI-generated audio and video to independent safety testing before public release.
WebArena: A Realistic Web Environment for Building Autonomous Agents
WebArena is a reproducible environment of four fully functional websites — e-commerce, a social forum, collaborative software development and content management — built for testing language-guided web agents. Its best GPT-4 agent completed 14.41% of tasks; humans completed 78.24%.[^1]
Stability AI releases SDXL 1.0
On 26 July 2023 Stability AI announced SDXL 1.0, its new flagship open image model, after a limited research-only release of SDXL 0.9; a one-step SDXL Turbo followed in November.
Anthropic, Google, Microsoft and OpenAI launch the Frontier Model Forum
An industry body formed on 26 July 2023 by Anthropic, Google, Microsoft and OpenAI to work on the safe and responsible development of frontier AI models, defined as large-scale models that exceed the capabilities of the most advanced existing models.
Google DeepMind introduces RT-2 vision-language-action model
On 28 July 2023, Google DeepMind introduced RT-2, a vision-language-action model that carries knowledge learned from web data over to robot control and roughly doubled generalisation to unseen scenarios compared with its predecessor.
ByteDance begins testing the Doubao chatbot and its Skylark model
In August 2023 a ByteDance subsidiary opened its Doubao chatbot for invitation testing, built on a model called Skylark that was among the first LLMs to pass China's algorithm registry.
Alibaba open-sources Qwen-7B, its first open-weight model
Alibaba's first open-weight Qwen releases: Qwen-7B appeared on 3 August 2023, followed by 14B, 1.8B and 72B models, giving outside developers downloadable models from small to large, with published training volumes, context limits and benchmark scores.
iFlytek releases SparkDesk V2.0 and iFlyCode
On 15 August 2023 iFlytek released SparkDesk V2.0, upgrading its large language model's coding and multimodal abilities and launching the iFlyCode programming assistant. The launch was followed by an 8.63% fall in the company's shares.
Unitree releases the H1 humanoid
On 15 August 2023, Unitree released the H1, its first general-purpose humanoid robot, a roughly 180 cm, 47 kg machine that the company described as China's first full-size general-purpose humanoid able to run.
AgiBot unveils the Yuanzheng A1
On 18 August 2023, AgiBot released the Yuanzheng A1, its first humanoid robot, at its first public product launch. The machine was pitched as a large-model-driven, embodied-intelligence body aimed first at manufacturing.
Meta releases Code Llama
On 24 August 2023 Meta released Code Llama, a family of code models built by further training Llama 2 on code. It came in 7B, 13B and 34B sizes, under the same community licence as Llama 2, and a 70B version followed in January 2024.
Hugging Face raises $235 million at a $4.5 billion valuation
On 24 August 2023 Hugging Face raised a $235 million Series D led by Salesforce Ventures, valuing the open-model hub at $4.5 billion, more than double its 2022 valuation, with chipmakers and cloud providers among the investors.
Alibaba releases the Qwen-VL vision-language models
In August 2023 Alibaba released the Qwen-VL vision-language models, built on the Qwen language model. The paper was posted to arXiv on 24 August 2023, and the code, demos and models were made public.
AI21 Labs closes $155 million Series C
On 31 August 2023, AI21 Labs announced a $155 million Series C round with Google and Nvidia among the new investors, bringing its total capital raised to $283 million at a reported $1.4 billion valuation.
ERNIE Bot opens to the public after regulatory approval
On 31 August 2023 Baidu opened its ERNIE Bot chatbot fully to the public through a website and mobile app after Chinese regulatory approval; more than a million users tried it within a day.
Baichuan releases the open Baichuan 2 models
On 6 September 2023 Baichuan released Baichuan 2, its next-generation open model family in 7B and 13B sizes, trained on 2.6 trillion tokens and published with intermediate checkpoints.
Anthropic launches the Claude Pro subscription
On 7 September 2023 Anthropic introduced Claude Pro, its first paid consumer plan: $20 a month in the US or GBP 18 in the UK, with five times the free tier's usage of Claude 2, priority access at busy times and early access to new features.
Tencent unveils its Hunyuan foundation model
On 7 September 2023 Tencent unveiled Hunyuan, its proprietary large foundation model, at a summit in Shenzhen. It was presented to enterprises, with access through Tencent Cloud APIs.
Microsoft releases phi-1.5
On 11 September 2023 Microsoft researchers posted the phi-1.5 technical report, a 1.3-billion-parameter model for natural-language common-sense reasoning, released openly under the MIT licence.
Stability AI launches Stable Audio
On 13 September 2023 Stability AI released Stable Audio, a web tool that generates music and sound from text prompts, built on a latent diffusion model trained with licensed content from the AudioSparx library.
Databricks raises $500 million Series I at a $43 billion valuation
On 14 September 2023 Databricks raised more than $500 million in a Series I round at a $43 billion valuation, led by funds advised by T. Rowe Price, about two months after closing its roughly $1.3 billion MosaicML acquisition [^1][^2].
Anthropic publishes its Responsible Scaling Policy and Long-Term Benefit Trust
On 19 September 2023 Anthropic published its Responsible Scaling Policy, a set of protocols for managing catastrophic risks from increasingly capable AI, and described the Long-Term Benefit Trust that oversees changes to it.
OpenAI announces DALL·E 3, built into ChatGPT
On 20 September 2023 OpenAI announced DALL·E 3, its third-generation image model, which was then built directly into ChatGPT; it was designed to refuse requests that imitate the style of living artists.
Amazon agrees to invest up to $4 billion in Anthropic
On 25 September 2023 Amazon agreed to invest up to $4 billion in Anthropic and take a minority stake. AWS became Anthropic's primary cloud provider and gave Anthropic access to Trainium and Inferentia chips.
Amazon commits up to $4 billion to Anthropic
On 25 September 2023 Anthropic announced that Amazon would invest up to $4 billion for a minority stake. Amazon Web Services became Anthropic's primary cloud provider for mission-critical workloads.
ChatGPT can now see, hear and speak
On 25 September 2023 OpenAI announced that ChatGPT could see, hear and speak, adding image input powered by GPT-4V and voice conversations; two days later it also gained web access to current information.
Preferred Networks releases its first large language model, PLaMo-13B
On 25 September 2023, Preferred Networks published PLaMo-13B, its first large language model and its first open-weight release, trained on 1.5 trillion tokens and released under Apache-2.0.
WGA and SAG-AFTRA deals set first AI protections for Hollywood writers and performers
In 2023, Hollywood writers and performers reached contracts with studios that set rules for artificial intelligence, including limits on AI-written material and consent requirements for digital replicas of performers.
Meta introduces the Meta AI assistant and AI characters
At Meta Connect on 27 September 2023 Meta introduced Meta AI, a conversational assistant in US beta on WhatsApp, Messenger and Instagram, along with 28 AI characters, 14 of them played by celebrities and influencers. It was rebuilt on Llama 3 in April 2024.
Mistral AI releases Mistral 7B, its first open-weight model
On 27 September 2023 Mistral AI released Mistral 7B, a 7.3-billion-parameter model with Apache 2.0 weights, less than five months after the company was founded. It was the start of the company's open-weight line.
Google invests in Anthropic
In October 2023 Google invested $500 million in Anthropic and committed $1.5 billion more over time, later agreeing to another $1 billion and, in 2025, a Google Cloud TPU deal worth tens of billions of dollars.
Moonshot AI launches the long-context Kimi chatbot
In October 2023 Moonshot AI released Kimi, a chatbot built around very long context, at first in closed beta; it opened to the public in November 2023, and by March 2024 it could take in two million Chinese characters at once.
A Cruise driverless car drags a pedestrian in San Francisco
On 2 October 2023 a driverless Cruise car in San Francisco struck a pedestrian and, according to NHTSA, dragged the person about 20 feet before stopping. Cruise paused driverless operations nationwide, recalled all 950 cars, and California's DMV then suspended its driverless permits.
SWE-bench: Can Language Models Resolve Real-World GitHub Issues?
A benchmark of 2,294 real software engineering problems taken from GitHub issues and the pull requests that resolved them, in which a model must edit a codebase to fix an issue; on release the best model solved under 2% of them.
Baidu launches ERNIE 4.0 at Baidu World 2023
On 17 October 2023 Baidu CEO Robin Li unveiled ERNIE 4.0 at Baidu World 2023 in Beijing, claiming it was "not inferior in any aspect to GPT-4" and announcing a paid ERNIE Bot plan built on it.
US tightens advanced chip export controls on China
On 17 October 2023 the US Bureau of Industry and Security issued three rules updating its October 2022 chip export controls, changing the chip parameters, adding licence exceptions and requirements, and expanding the Entity List.
Adept releases the encoder-free multimodal model Fuyu-8B
On 17 October 2023 Adept released Fuyu-8B, an encoder-free multimodal model whose weights and example code went to Hugging Face the same day, positioned as part of a copilot for knowledge workers.
China issues guidance on humanoid robot innovation
On 20 October 2023 China's Ministry of Industry and Information Technology issued guidance on humanoid robots, setting targets for 2025 and 2027 and five groups of tasks covering technology, products, scenarios, ecosystem and support.
iFlytek releases Spark V3.0 and claims it beats GPT-3.5 in Chinese
On 24 October 2023 iFlytek released Spark V3.0 and claimed it beat GPT-3.5 in Chinese across six areas, unveiled a Huawei-built computing platform the same day, and saw its stock fall to the daily limit.
California suspends Cruise's driverless permits
On 24 October 2023 California's DMV suspended Cruise's driverless testing and deployment permits with immediate effect, and the company then stopped driverless operations in all markets.
Biden signs Executive Order 14110 on AI
On 30 October 2023, President Biden signed Executive Order 14110, a federal directive on AI safety, security and trustworthy use. It set principles and assigned more than 100 actions across government.
First global AI Safety Summit adopts the Bletchley Declaration
On 1–2 November 2023 the UK hosted the first global AI Safety Summit at Bletchley Park, where 28 countries including the US, China and the EU agreed the Bletchley Declaration on the risks of frontier AI.
01.AI releases the open Yi-34B and Yi-6B models
On 2 November 2023, 01.AI released the open Yi-6B and Yi-34B base models. Yi-34B ranked first among open pre-trained models on two public leaderboards, ahead of Meta's Llama 2.
DeepSeek releases its first model, DeepSeek Coder
On 2 November 2023 DeepSeek released its first model, DeepSeek Coder, a set of open-source code models from 1.3B to 33B parameters trained on 2 trillion tokens each.
xAI announces Grok, its first chatbot and model
On 3 November 2023 xAI announced Grok, its first chatbot, powered by the new Grok-1 model, and offered early access to a limited number of US users.
OpenAI DevDay 2023: GPT-4 Turbo and custom GPTs
On 6 November 2023 OpenAI held its first developer conference, DevDay, announcing GPT-4 Turbo with a 128K context window, custom GPTs built with a GPT Builder, and lower prices across its platform.
NVIDIA announces H200 at SC23
On 13 November 2023 NVIDIA announced the HGX H200 at the SC23 conference in Denver, an addition to its Hopper line that twelve server partners and four major cloud providers were named to offer, with shipments expected from the second quarter of 2024.
GraphCast weather model is published in Science
On 14 November 2023, Google DeepMind announced GraphCast; its Science paper reported beating ECMWF HRES on most test variables, and a live version forecast Hurricane Lee's landfall earlier.
Bing Chat is renamed Microsoft Copilot
On 15 November 2023 Microsoft renamed Bing Chat and Bing Chat Enterprise to Copilot, giving the assistant its own site and a commercial version with data protection.
OpenAI's board fires Sam Altman, who returns days later
In November 2023 OpenAI's board removed CEO Sam Altman, citing a lack of candour; after an employee revolt and a Microsoft offer, he was reinstated within days under a new board.
GPQA: A Graduate-Level Google-Proof Q&A Benchmark
GPQA is a benchmark of 448 expert-written multiple-choice questions in biology, physics and chemistry, built so that skilled non-experts cannot look the answers up, and hard enough that the strongest GPT-4 baseline scored 39% while domain experts reached 65%.
Stability AI releases Stable Video Diffusion
On 21 November 2023 Stability AI released Stable Video Diffusion, its first generative video foundation model, as a research preview of two image-to-video models with 14 and 25 frames.
Anthropic releases Claude 2.1 with a 200K context window
On 21 November 2023 Anthropic released Claude 2.1, doubling the context window to 200,000 tokens. It also added a beta tool use feature and system prompts for setting Claude's role and style.
GAIA: a benchmark for General AI Assistants
GAIA is a benchmark of real-world questions that are easy for humans but hard for AI assistants, testing reasoning, multimodality, web browsing, and tool use. It compares AI progress against average human performance.
AWS introduces the Amazon Q assistant
On 28 November 2023, at re:Invent, AWS introduced Amazon Q, a generative AI assistant for work, in preview. Answers respect users' permissions and cite sources, and Amazon Q Business became generally available on 30 April 2024.
Amazon makes Titan text models available and previews Titan Image Generator
On 29 November 2023, during AWS re:Invent, Amazon made the Titan Text Express and Titan Text Lite models generally available in Amazon Bedrock and previewed Titan Image Generator, which adds an invisible watermark to every image.
DeepSeek releases DeepSeek LLM 7B and 67B
On 29 November 2023 DeepSeek released DeepSeek LLM in 7B and 67B sizes, each with base and chat versions, trained on 2 trillion tokens.
Google DeepMind's GNoME finds 2.2 million new crystals
On 29 November 2023 Google DeepMind announced GNoME, a deep-learning system that predicted 2.2 million new crystal structures, 380,000 of them among the most stable; hundreds of the predictions were independently made by outside researchers.
Scaling deep learning for materials discovery
A 2023 Nature paper from researchers at Google DeepMind presenting a scaled-up deep-learning approach to predicting stable crystal structures, screening 2.2 million candidates and reporting a large jump in stability-prediction accuracy.
LAION-5B taken offline over CSAM links and re-released as Re-LAION-5B
A December 2023 Stanford Internet Observatory report found links to illegal images in LAION-5B. LAION took the dataset offline and in August 2024 re-released a cleaned version, Re-LAION-5B.
Mamba: Linear-Time Sequence Modeling with Selective State Spaces
Mamba is a sequence-model architecture introduced by Albert Gu and Tri Dao that replaces attention with selective state space models, giving linear-time scaling, faster inference, and language-model quality competitive with Transformers twice its size.
Google DeepMind introduces Gemini
On 6 December 2023 Google announced Gemini 1.0, a model family built to learn text, images, audio and video together from the start. Google called its largest version its most capable model yet.
EU co-legislators reach a provisional deal on the AI Act
On 8 December 2023 negotiators for the Council and the European Parliament reached a provisional agreement on the AI Act, the file the Council described as the first rules for AI in the world; formal adoption followed in May 2024.
Mistral AI releases Mixtral 8x7B, an open mixture-of-experts model
On 11 December 2023 Mistral AI released Mixtral 8x7B, an open-weight mixture-of-experts model that activates only part of its parameters for each word. It aimed to pair a large total capacity with lower compute per token.
Mistral AI raises about €385 million in its Series A
On 11 December 2023 Mistral AI announced a €385 million Series A — its second round in seven months — the same day it released the Mixtral 8x7B model and opened its API platform.
Microsoft releases Phi-2
On 12 December 2023 Microsoft Research announced Phi-2, a 2.7-billion-parameter model it pitched as a playground for researchers, released under the MIT licence.
Tesla reveals Optimus Gen 2
On 12 December 2023 Tesla posted an unannounced video of Optimus Gen 2, a lighter, faster humanoid with Tesla-designed actuators and sensors, an 11-degree-of-freedom hand and tactile sensing.
FunSearch produces new results on open mathematics problems
FunSearch, published in Nature in December 2023, used large language models for program search rather than direct solution search, and found results better than the previous best on the cap set problem and online bin packing.
OpenAI publishes its Preparedness Framework
A beta document published by OpenAI in December 2023 that sets out processes to track, evaluate, forecast and protect against catastrophic risks from powerful models, rating models low to critical and barring deployment above medium.
US authors sue OpenAI and Microsoft over training data
Eleven US authors and a later class action sued OpenAI and Microsoft, alleging the companies copied books without permission to train AI chatbots, part of a wider wave of copyright litigation.
Midjourney releases V6
On 21 December 2023 Midjourney released version 6 of its image model in alpha, trained from scratch over nine months and better at rendering text. V6.1 followed on 30 July 2024 with faster, more coherent output and became the default model.
The New York Times sues OpenAI and Microsoft over copyright
On 27 December 2023 The New York Times sued OpenAI and Microsoft in Manhattan federal court, saying millions of its articles were used without permission to build AI products including ChatGPT — the first such suit by a major US media organisation.
UBTECH lists on the Hong Kong Stock Exchange
UBTECH, a Chinese robot maker founded in 2012, listed on the Hong Kong Stock Exchange main board on 29 December 2023 at HK$90 per share; the company calls itself the first humanoid robot company on the exchange.
2024
Andrew Barto and Richard Sutton receive the 2024 ACM Turing Award
In 2024 the ACM gave its A.M. Turing Award to Andrew Barto and Richard Sutton for building the conceptual and algorithmic foundations of reinforcement learning, work that runs from 1980s papers to a standard textbook and on to deep reinforcement learning.
Mixtral of Experts
Mixtral 8x7B is a sparse mixture-of-experts language model from Mistral AI. A router sends each token to two of eight experts per layer, so the model holds 47B parameters but uses 13B during inference.
DeepSeek introduces the DeepSeekMoE architecture
On 9 January 2024 DeepSeek released DeepSeekMoE, a mixture-of-experts architecture built around fine-grained experts and shared experts, with an open-weights 16B model.
OpenAI opens the GPT Store
On 10 January 2024 OpenAI opened the GPT Store, a marketplace for custom versions of ChatGPT. More than three million GPTs had already been built by the time it launched, after the opening slipped from its originally planned date.
Anthropic and Redwood Research publish 'Sleeper Agents'
A January 2024 paper by Anthropic and Redwood Research showed that deliberately planted backdoor behavior in language models could survive supervised fine-tuning, reinforcement learning and adversarial training, and that adversarial training could even teach a model to hide it.
IMF analysis finds almost 40% of global employment exposed to AI
An IMF staff analysis published in January 2024 estimated that almost 40 percent of global employment is exposed to AI, with effects uneven across income levels and a risk of widening inequality.
AlphaGeometry solves olympiad geometry problems
On 17 January 2024 Google DeepMind announced AlphaGeometry, a system that pairs a language model with a symbolic deduction engine and solved 25 of 30 olympiad geometry problems in a benchmark, close to the 25.9 averaged by human gold medallists.
Figure and BMW sign a commercial deployment agreement
On 18 January 2024, Figure announced a commercial agreement with BMW Manufacturing to deploy general-purpose robots in car production, starting with a few tasks at BMW's Spartanburg plant and expanding only if milestones were met.
AI-cloned Biden voice robocalls hit New Hampshire primary voters
Two days before New Hampshire's 2024 primary, voters received robocalls using an AI-generated imitation of President Joe Biden's voice telling them to skip the vote; the case brought the FCC's first generative-AI fine and a ban on AI voices in robocalls.
01.AI open-sources the Yi-VL vision-language models
On 23 January 2024, 01.AI released the open vision-language models Yi-VL-34B and Yi-VL-6B. Yi-VL-34B ranked first among open models on the MMMU and CMMMU benchmarks.
Baichuan releases Baichuan 3, a hundred-billion-parameter model
On 29 January 2024 Baichuan released Baichuan 3, a closed model with more than one hundred billion parameters, and reported its own benchmark results against GPT-4.
iFlytek releases Spark V3.5, trained on domestic compute
On 30 January 2024 iFlytek released Spark V3.5, which it described as the first Spark model trained on the fully domestic FlyingStar One platform and which it compared with GPT-4 Turbo across several benchmarks.
Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Research
Dolma is an openly released English corpus of three trillion tokens, built from web text, papers, code, books, social media and Wikipedia, with a paper documenting both the dataset and the pipeline used to curate it.
Cohere For AI releases Aya, an open model for 101 languages
In February 2024 Cohere For AI released Aya, an instruction-tuned model covering 101 languages. It has 13 billion parameters, uses an Apache 2.0 licence and beats mT0 and BLOOMZ on most tasks, according to its authors.
Alibaba leads a US$1 billion round in Moonshot AI
In February 2024 Alibaba led a US$1 billion funding round in Moonshot AI at a US$2.5 billion valuation, roughly eight times the startup's earlier valuation and, by reporting at the time, the largest single round for a Chinese AI startup.
Arup finance worker pays $25 million after deepfake video call
In February 2024 Hong Kong police revealed that a finance worker at a multinational had been tricked into paying HK$200 million (about US$25.6 million) after a video call in which every other participant was a deepfake. The employer was later confirmed to be the engineering firm Arup.
Ai2 releases OLMo, a fully open language model
On 1 February 2024, the Allen Institute for AI released OLMo, a language model released together with its pretraining data, training code, logs and more than 500 checkpoints per model, all under the Apache 2.0 licence.
DeepSeekMath introduces the GRPO reinforcement-learning algorithm
A February 2024 paper introduced DeepSeekMath 7B, a math-focused language model, and with it Group Relative Policy Optimization, a reinforcement-learning method that drops PPO's critic model.
Alibaba releases the Qwen1.5 open model family
On 4 and 5 February 2024 Alibaba released the Qwen1.5 open model family, with base and chat models from 0.5B to 72B parameters and a 32K context window across the lineup.
SenseTime releases SenseNova 4.0
On 8 February 2024 SenseTime announced SenseNova 4.0, including SenseChat V4 in 4K, 32K and 128K context versions, specialised models for data analysis, medicine and vision, a tool-calling API, and a HumanEval score above GPT-4's.
Bard is renamed Gemini and Gemini Advanced launches
On 8 February 2024 Google renamed its assistant Bard to Gemini, matching the name of its model family, and launched Gemini Advanced on the Ultra 1.0 model through a paid Google One plan in more than 150 countries.
FCC rules AI-generated voices in robocalls illegal
In February 2024 the FCC ruled that AI-generated voices in robocalls count as "artificial" under the Telephone Consumer Protection Act, making them subject to the same restrictions as prerecorded calls — a response to a deepfake Biden robocall sent before the New Hampshire primary.
ChatGPT starts remembering users across conversations
In 2024 OpenAI gave ChatGPT a memory, letting it carry information about a user from one conversation to the next, with switches to turn the feature off and a temporary chat mode that leaves no memory.
Air Canada held liable for its chatbot's wrong bereavement-fare advice
In 2024 a Canadian tribunal held Air Canada responsible for wrong bereavement-fare advice given by the chatbot on its website, rejecting the airline's claim that the chatbot was a separate legal entity.
OpenAI previews Sora, its text-to-video model
On 15 February 2024 OpenAI previewed Sora, a text-to-video model that it said could generate clips up to one minute long; the model reached ChatGPT Plus and Pro users in the US and Canada on 9 December 2024.
Google announces Gemini 1.5 Pro with a 1-million-token context window
Announced on 15 February 2024, Gemini 1.5 Pro was a mixture-of-experts model offering a 128,000-token context window by default and up to 1 million tokens in private preview — far more material taken in at once.
Google releases Gemma open models
On 21 February 2024 Google released Gemma, a family of open models in 2B and 7B sizes, with pre-trained and instruction-tuned versions, weights distributed through Kaggle and Hugging Face and terms permitting commercial use by organisations of any size.
Stability AI previews Stable Diffusion 3
On 22 February 2024 Stability AI announced an early preview of Stable Diffusion 3, a text-to-image suite from 800M to 8B parameters that combines a diffusion transformer with flow matching.
Google pauses Gemini's people image generation after historically inaccurate images
On 22 February 2024 Google paused Gemini's generation of images of people after users shared historically inaccurate pictures; Google blamed diversity tuning applied where no range was appropriate, and over-cautious refusals.
Mistral AI launches Mistral Large and a partnership with Microsoft
On 26 February 2024 Mistral AI released Mistral Large and Mistral Small, and announced a multi-year partnership with Microsoft that made Mistral Large available first on Azure AI. Le Chat, its assistant, launched in beta the same day.
Mistral AI launches the Le Chat assistant
On 26 February 2024, Mistral AI's API changelog announced Le Chat, a beta chat interface for talking to its models directly, released alongside the Mistral Large and Mistral Small models and new function calling and JSON mode features.
BigCode and NVIDIA release StarCoder2 and The Stack v2
On 28 February 2024 BigCode and NVIDIA released StarCoder2 in 3B, 7B and 15B sizes. The new training set, The Stack v2, is 67.5 TB and is built from the Software Heritage archive.
Figure partners with OpenAI on robot AI models
On 29 February 2024, humanoid robot maker Figure announced a US$675 million Series B at a US$2.6 billion valuation and a collaboration with OpenAI to develop AI models for its robots.
Microsoft hires Inflection AI's founders and staff in a $650 million licensing deal
In March 2024 Microsoft hired nearly all of Inflection AI's staff, including co-founders Mustafa Suleyman and Karén Simonyan, and paid $650 million to license its technology — a deal the UK competition authority later examined as a merger.
Kimi extends its context window to 2 million Chinese characters
In March 2024 Moonshot AI opened a closed beta of its Kimi assistant with a 2 million Chinese character context window, roughly ten times what the chatbot handled at launch, and gave it to invited users first.[^1][^2][^3]
Anthropic introduces the Claude 3 model family
On 4 March 2024 Anthropic announced the Claude 3 family: Haiku, Sonnet and Opus, in ascending order of capability. All three added vision and launched with a 200K-token context window.
Amazon buys Talen's Cumulus data center campus next to the Susquehanna nuclear plant
In March 2024 Talen Energy sold its Cumulus data center campus beside the Susquehanna nuclear plant to Amazon Web Services for $650 million, with 960 MW of plant output going to AWS; a 2025 deal expanded it to 1,920 MW through 2042.
Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context
A report from Google's Gemini Team introducing the Gemini 1.5 family, multimodal models built around very long context, with Gemini 1.5 Pro recalling information across up to 10 million tokens.[^1]
Cohere releases Command R for enterprise RAG and tool use
On 11 March 2024 Cohere released Command R, a 35-billion-parameter model built for retrieval-augmented generation and tool use. Its weights were released for research use on Hugging Face.
LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Code
A benchmark for code LLMs that keeps collecting fresh contest problems from LeetCode, AtCoder and CodeForces, so that scores reflect ability rather than memorised test data, and that also tests code repair and execution reasoning.
Claude 3 Haiku becomes available
On 13 March 2024 Anthropic made Claude 3 Haiku available, calling it the fastest and most affordable model in its intelligence class. It was priced at a one-to-five input-to-output ratio and accepted image input.
European Parliament and Council adopt the AI Act
The European Parliament approved the AI Act on 13 March 2024 and the Council of the EU formally adopted it on 21 May 2024; the Council described the act as the first worldwide rules on AI.
Apptronik and Mercedes-Benz announce an Apollo pilot
On 15 March 2024 Apptronik announced a pilot with Mercedes-Benz to test Apollo humanoid robots in its manufacturing facilities — Apptronik's first publicly announced commercial deployment of Apollo.
xAI releases Grok-1's weights under Apache 2.0
On 17 March 2024 xAI released the base weights and architecture of Grok-1, a 314-billion-parameter mixture-of-experts model, under the Apache 2.0 license.
NVIDIA announces Project GR00T
In March 2024 NVIDIA announced Project GR00T, a general-purpose foundation model for humanoid robots. A year later it released Isaac GR00T N1, which it presented as a fully customizable humanoid model.
NVIDIA announces the Blackwell platform
On 18 March 2024 NVIDIA announced Blackwell, the GPU platform succeeding Hopper, claiming it can run real-time generative AI on trillion-parameter models at up to 25 times lower cost and energy.
Mustafa Suleyman joins Microsoft to lead the new Microsoft AI
On 19 March 2024 Microsoft announced that Mustafa Suleyman and Karén Simonyan were joining to form Microsoft AI, which took over Copilot and the consumer AI products.
Sakana AI merges open models by evolutionary search to build a Japanese LLM
On 19 March 2024, Sakana AI posted a paper on using evolutionary search to merge open models without extra training data or compute, and released a 7B Japanese model that also handles mathematics as an experiment.
Emad Mostaque resigns as Stability AI CEO
On 23 March 2024, Emad Mostaque stepped down as CEO and board director of Stability AI; the company named interim co-CEOs and began searching for a permanent leader.
StepFun debuts with the Step-1, Step-1V and Step-2 preview models
On 23 March 2024 StepFun made its public debut in Shanghai, announcing Step-1, Step-1V and a preview of the trillion-parameter Step-2, plus two consumer apps.
Databricks releases the 132B mixture-of-experts model DBRX
Databricks released DBRX, a mixture-of-experts model shipped in base and instruct versions, on 27 March 2024. The company framed it as competition for Mixtral and Grok-1 rather than closed frontier systems, and reported it ahead of LLaMA2-70B Chat on every benchmark it published.
xAI announces Grok-1.5 with a 128K context
On 28 March 2024 xAI announced Grok-1.5, with better reasoning and a 128,000-token context window. About two weeks later it previewed Grok-1.5V, its first multimodal model.
AI21 Labs releases Jamba, a hybrid Mamba-Transformer model
On 28 March 2024 AI21 Labs released Jamba, an open-weights model that combines Mamba state-space layers with Transformer attention. It has 52 billion total parameters and a 256k-token context.
Langboat open-sources the Mengzi3 model family
On 29 March 2024 Langboat released Mengzi3, an open-weight language model family in 8B and 13B sizes under Apache-2.0, with base weights first and a chat version following later.
China publishes consolidated filings for generative AI services
On 2 April 2024 China's cyberspace administration began publishing which generative AI services had completed filing, and required launched apps to display the model name and filing number they use.
Stability AI releases Stable Audio 2.0
On 3 April 2024 Stability AI released Stable Audio 2.0, a model that generates stereo tracks of up to three minutes, adds audio-to-audio transformation, and was trained on more than 800,000 licensed files.
Cohere launches Command R+, first on Microsoft Azure
On 4 April 2024 Cohere introduced Command R+, a retrieval-augmented model for enterprise workloads. It was first offered on Microsoft Azure, where Microsoft said it works with Azure AI Search and Azure AI Content Safety.
OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments
OSWorld is a benchmark and environment for testing multimodal agents on 369 real computer tasks in Ubuntu, Windows and macOS; humans complete over 72.36% of the tasks, while the best model reaches only 12.24%.
Boston Dynamics retires hydraulic Atlas and unveils electric Atlas
In April 2024 Boston Dynamics retired its hydraulic Atlas and introduced a fully electric humanoid, promising more strength, dexterity and range of motion, with Hyundai's factories as the first testing ground.
Mistral AI releases Mixtral 8x22B
On 17 April 2024 Mistral AI announced Mixtral 8x22B, a sparse mixture-of-experts language model with 141 billion total but only 39 billion active parameters, released under Apache 2.0 with a 64K context window.
MiniMax launches the abab 6.5 mixture-of-experts series
On 17 April 2024, MiniMax released the abab 6.5 series, a family of mixture-of-experts language models reported to support a 200,000-token context. It extended the company's proprietary abab line, which began with abab1 in 2022.
Meta releases Llama 3
On 18 April 2024 Meta released Llama 3 in 8B and 70B sizes, both pretrained and instruction-tuned. The models were pretrained on over 15 trillion tokens and made available through major cloud providers and Hugging Face under the Meta Llama 3 Community License.
Apple releases the OpenELM open language models
On 22 April 2024 Apple researchers posted the OpenELM paper to arXiv and released eight small language models, with training logs, checkpoints, configurations and code for running the models on Apple devices.
Microsoft launches Phi-3-mini
On 23 April 2024 Microsoft introduced the Phi-3 family of small language models, starting with Phi-3-mini, released on Azure AI Studio, Hugging Face and Ollama.
SenseTime launches SenseNova 5.0, benchmarking against GPT-4 Turbo
On 23 April 2024 SenseTime released SenseNova 5.0 and benchmarked it against GPT-4 Turbo, and its shares surged more than 30% before trading was suspended.
Snowflake open-sources the 480B hybrid model Arctic
On 24 April 2024 Snowflake announced Arctic, a large language model aimed squarely at enterprise work such as SQL generation, coding and instruction following, and released it with training recipes as well as weights.
NVIDIA agrees to acquire GPU orchestration software provider Run:ai
In April 2024 NVIDIA agreed to acquire Run:ai, an Israeli maker of Kubernetes-based GPU orchestration software, reported at about $700 million. The European Commission approved the deal in December and it closed on 30 December 2024, with the software to be open-sourced.
GitHub Copilot Workspace enters technical preview
On 29 April 2024 GitHub opened the technical preview of Copilot Workspace, a task-centric environment where a repository, issue or pull request becomes an editable plan and code, with the developer keeping control.
Scale AI closes a $1 billion Series F at a $13.8 billion valuation
In May 2024 Scale AI closed a $1 billion financing at a $13.8 billion valuation, led by existing investor Accel, with participation from existing backers such as NVIDIA and new ones including Amazon, Meta and Intel Capital.
Anthropic launches the Claude Team plan and iOS app
On 1 May 2024 Anthropic announced a Claude Team plan at $30 per user per month with a five-seat minimum, plus a free Claude iOS app that syncs chat history with the web.[^1]
CoreWeave raises a $1.1 billion Series C at a $19 billion valuation
On 1 May 2024 CoreWeave announced a $1.1 billion Series C round led by Coatue at a $19 billion valuation, up from $7 billion a few months earlier, and weeks later closed a $7.5 billion debt facility led by Blackstone.
DeepSeek releases DeepSeek-V2 with Multi-head Latent Attention
On 7 May 2024 DeepSeek published the DeepSeek-V2 technical report, a 236B mixture-of-experts model with 21B activated parameters per token and Multi-head Latent Attention.
Google DeepMind and Isomorphic Labs unveil AlphaFold 3
On 8 May 2024, Google DeepMind and Isomorphic Labs announced AlphaFold 3, a model that predicts the structures and interactions of proteins, DNA, RNA and ligands, and that Google reported as the first AI system to outperform physics-based tools for biomolecular structure prediction.
Accurate structure prediction of biomolecular interactions with AlphaFold 3
AlphaFold 3, published in Nature in May 2024, extends structure prediction to biomolecular interactions, reporting greater accuracy than docking tools for protein–ligand binding and than AlphaFold-Multimer v.2.3 for antibody–antigen complexes.
Unitree introduces the G1 humanoid robot
On 13 May 2024, Hangzhou-based Unitree Robotics launched the G1 humanoid robot at a starting price of 99,000 yuan, well below the 150,000–200,000 yuan that China Daily reported for many competing machines.
OpenAI introduces GPT-4o
On 13 May 2024 OpenAI announced GPT-4o ("GPT-4 Omni"), its new flagship model, which it said can reason across audio, vision and text in real time. Developers could use it through the API the same day.
01.AI open-sources Yi-1.5 under Apache 2.0
On 13 May 2024 01.AI open-sourced the Yi-1.5 series in 34B, 9B and 6B sizes under Apache 2.0, as part of its first-anniversary event that also launched the closed Yi-Large model.
01.AI launches Yi-Large and the Wanzhi app
On 13 May 2024 01.AI launched the closed Yi-Large model and the free Wanzhi app at its first-anniversary event, and opened its API platform worldwide the same day.
Tencent open-sources the Hunyuan-DiT text-to-image model
On 14 May 2024 Tencent posted the technical report for Hunyuan-DiT, a 1.5-billion-parameter text-to-image model, and publicly released its code and pretrained models.
Google DeepMind shows Project Astra at I/O
At Google I/O on 14 May 2024, Demis Hassabis presented Project Astra, Google DeepMind's vision for future AI assistants: a system that continuously encodes video and speech into a cacheable timeline of events.
Google launches AI Overviews in Search
At Google I/O in May 2024, Google began rolling out AI Overviews, Gemini-generated summaries at the top of search results, to everyone in the US; the feature later reached over 100 countries and drew criticism for erroneous answers.
Google unveils Imagen 3 and Veo at I/O
On 14 May 2024 at Google I/O, Google announced Imagen 3, its highest-quality text-to-image model at the time, and Veo, a video model producing 1080p clips over a minute long. Both opened in private preview.
Google introduces Gemini 1.5 Flash at I/O 2024
On 14 May 2024, at Google I/O, Google introduced Gemini 1.5 Flash, a lighter and faster model distilled from Gemini 1.5 Pro, alongside a 2-million-token context window for 1.5 Pro, Project Astra and Gemma 2.
Volcano Engine launches the Doubao model family and cuts LLM prices
On 15 May 2024 Volcano Engine, ByteDance's cloud unit, launched the Doubao model family and priced its main model far below rivals, starting a price war in China's enterprise AI market.
Google DeepMind introduces its Frontier Safety Framework
On 17 May 2024 Google DeepMind published a framework of protocols for identifying AI capabilities that could cause severe harm, evaluating models against defined Critical Capability Levels, and applying mitigations when a model reaches them.
OpenAI dissolves the Superalignment team after Sutskever and Leike exit
In May 2024 OpenAI's chief scientist Ilya Sutskever left the company, followed days later by Superalignment co-leader Jan Leike; OpenAI then disbanded the team, and Leike publicly said safety had "taken a backseat to shiny products".
AI Seoul Summit adopts the Seoul Declaration
On 21 May 2024, ten countries and the European Union signed the Seoul Declaration at the AI Seoul Summit, one of three outcome documents that tied government statements to the Frontier AI Safety Commitments and the Hiroshima Process International Code of Conduct.
Anthropic publishes 'Mapping the Mind of a Large Language Model'
On May 21, 2024, Anthropic published research that used dictionary learning to map millions of concepts, called features, inside Claude 3 Sonnet — including one that responds to the Golden Gate Bridge.
Baichuan launches Baichuan 4 and the Baixiaoying assistant
On 22 May 2024 Baichuan unveiled Baichuan 4 and its assistant Baixiaoying in Beijing, presenting the assistant as a step toward an AI super app.
iFlytek makes Spark Lite API free in China's model price war
On 22 May 2024 iFlytek made its Spark Lite API permanently free and opened Spark API capabilities for free, while pricing its top-tier Spark3.5 Max at RMB 0.21 per 10,000 tokens, in the same weeks that ByteDance, Alibaba Cloud, Baidu and Tencent cut model prices.
Cohere For AI releases Aya 23 open-weight models
On 23 May 2024 Cohere For AI published Aya 23, a family of multilingual models covering 23 languages. Its 8B and 35B open-weight models were built on Cohere's Command family and carry a non-commercial licence.
Anthropic demos 'Golden Gate Claude', a model obsessed with the Golden Gate Bridge
A 24-hour public demo in which Anthropic clamped a single internal feature — the Golden Gate Bridge — at ten times its maximum activation in Claude 3 Sonnet, so that the model steered everyday answers back to the bridge.
Mistral AI releases Codestral, its first code model
On 29 May 2024 Mistral AI released Codestral, a 22-billion-parameter open-weight code model trained on more than 80 programming languages, with a 32K context window and weights under a non-production licence.
Tencent launches the Yuanbao AI assistant
Tencent released Yuanbao, a consumer AI assistant built on its own Hunyuan model, on 30 May 2024, offering search, chat, document handling and image generation across mobile, desktop and the web.
Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality
Tri Dao and Albert Gu's paper introduces state space duality, a framework showing that structured state space models and attention variants are closely related, and uses it to derive Mamba-2, whose core layer runs 2 to 8 times faster than Mamba's.
Amazon hires Adept's co-founders and licenses its agent technology
In June 2024 Amazon hired Adept's co-founders and some of its team and licensed the startup's agent technology, multimodal models and datasets, while Adept refocused on solutions that enable agentic AI under a new CEO.
MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark
MMLU-Pro is a 2024 benchmark that makes the widely used MMLU knowledge test harder: its questions are reasoning-focused and each has ten answer options instead of four.
Nvidia's market value breaks $3 trillion
On 5 June 2024 Nvidia's market value passed $3 trillion, making it the third company to reach that level after Microsoft and Apple; two weeks later it became the world's most valuable company, driven by demand for data-center AI chips.
Kuaishou launches the Kling video generation model
On 6 June 2024, Kuaishou launched Kling, a text-to-video model that generates up to two-minute 1080p clips, and opened testing first in China and later worldwide. Within a month, users had generated seven million videos.
Alibaba releases Qwen2, moving most sizes to Apache 2.0
On 6 and 7 June 2024 Alibaba released Qwen2 on Hugging Face and ModelScope. Most sizes moved to Apache 2.0, and the 72B model kept the Qianwen License.
Apple announces Apple Intelligence at WWDC 2024
On 10 June 2024, at WWDC, Apple announced Apple Intelligence for iOS 18, iPadOS 18 and macOS Sequoia. It introduced Private Cloud Compute and a free ChatGPT integration powered by GPT-4o.
Mistral AI raises €600 million at a €5.8 billion valuation
On 11 June 2024, Mistral AI closed a €600 million Series B round led by existing investor General Catalyst, valuing the company at €5.8 billion. Lightspeed, Andreessen Horowitz, Cisco Systems and Nvidia also took part.
Luma AI releases the Dream Machine video generator
On 12 June 2024 Luma Labs released Dream Machine, a free text-to-video generator that turned a written prompt into a realistic five-second clip in about two minutes, and quickly added clip extension, in-video editing and plans for APIs.
Stability AI releases open weights for Stable Diffusion 3 Medium
On 12 June 2024 Stability AI released the weights of Stable Diffusion 3 Medium, a 2-billion-parameter diffusion transformer, on Hugging Face under a non-commercial license.
OpenVLA released as open-source VLA
OpenVLA is an open-source vision-language-action model released in June 2024: trained on 970,000 robot manipulation trajectories and reported to beat a far larger prior model, with checkpoints, notebooks and code made public.
NVIDIA releases Nemotron-4 340B for synthetic data generation
On 14 June 2024 NVIDIA released Nemotron-4 340B, a family of base, instruct and reward models built for generating synthetic training data. It published the models under a permissive license, and the reward model topped RewardBench.
Runway announces Gen-3 Alpha
On 17 June 2024 Runway announced Gen-3 Alpha, a video generation model trained jointly on video and images, with finer temporal control and C2PA provenance metadata. It opened to paying subscribers on 1 July 2024.
DeepSeek releases DeepSeek-Coder-V2
On 17 June 2024 DeepSeek released DeepSeek-Coder-V2, an open-source mixture-of-experts code model trained on about 6 trillion additional tokens, with a 128K context window.
τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains
τ-bench is a benchmark, submitted to arXiv in June 2024, that tests language agents in dynamic conversations with a simulated user, with domain-specific API tools and policy rules, and scores them by database state rather than by the text of their replies.
Ilya Sutskever launches Safe Superintelligence Inc.
On 19 June 2024 Ilya Sutskever announced Safe Superintelligence Inc. on X, a company with one stated goal: building safe superintelligence. It raised $1 billion in September 2024.
Anthropic releases Claude 3.5 Sonnet and Artifacts
On 20 June 2024 Anthropic released Claude 3.5 Sonnet, the first model of the 3.5 family, and launched Artifacts on claude.ai. It cost the same as Claude 3 Sonnet and ran twice as fast as Claude 3 Opus.
Record labels sue Suno and Udio over AI music training
In June 2024 the RIAA announced that major record companies had sued Suno and Udio, two AI music generation services, alleging that building them required copying decades of copyrighted sound recordings.
Stability AI names Prem Akkaraju CEO and raises new funding
On 25 June 2024 Stability AI appointed Prem Akkaraju, former CEO of Weta Digital, as chief executive and named former Facebook president Sean Parker executive chairman, alongside new investment from Greycroft, Coatue and others; Bloomberg reported the round at $80 million.
Claude gains Projects
On 25 June 2024 Anthropic launched Projects on claude.ai for Pro and Team customers. Each project has a 200K-token context window and can hold documents, code and custom instructions.
EvolutionaryScale releases ESM3 and the esmGFP result
On 25 June 2024 EvolutionaryScale published ESM3, a protein language model trained at large scale, with a paper on the esmGFP generation experiment and an open checkpoint built on 2.78 billion natural protein sequences.
The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale
FineWeb is a 15-trillion-token pretraining dataset built from 96 Common Crawl snapshots and released in 2024 with its curation code and ablation models; its educational subset, FineWeb-Edu, improves results on knowledge- and reasoning-heavy benchmarks.
iFlytek releases Spark V4.0
On 27 June 2024 iFlytek released Spark V4.0, a large model it said was trained on a domestic 10,000-card cluster and beat GPT-4 Turbo on most measures while trailing on coding and multimodal work.
GXO and Agility sign the first commercial humanoid RaaS agreement
In June 2024 GXO Logistics and Agility Robotics announced a multi-year agreement to deploy Digit humanoid robots in GXO's logistics operations, which Agility called the industry's first commercial and first Robots-as-a-Service humanoid deployment.
Baidu launches ERNIE 4.0 Turbo as ERNIE Bot reaches 300 million users
On 28 June 2024 Baidu announced ERNIE 4.0 Turbo, describing it as faster and cheaper than ERNIE 4.0, and said ERNIE Bot had reached 300 million users.
StepFun releases Step-2, Step-1.5V and Step-1X at WAIC
At WAIC, which opened on 5 July 2024, StepFun released the official Step-2 language model, Step-1.5V and the Step-1X image generator.
SenseTime unveils SenseNova 5.5 and real-time SenseNova 5o at WAIC
At WAIC in July 2024 SenseTime unveiled SenseNova 5.5 and the real-time SenseNova 5o model, saying it beat GPT-4o on five of eight key metrics.
Kuaishou open-sources the Kolors text-to-image model
On 6 July 2024 Kuaishou released Kolors, a text-to-image model published as open source under Apache-2.0 with weights open for research, presented at a WAIC 2024 forum alongside KwaiYii and Kling.
SoftBank acquires Graphcore
In July 2024 SoftBank Group acquired Graphcore, which became a wholly owned subsidiary and kept its name and offices; the deal was recorded at about US$500 million, subject to UK government review.
FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision
FlashAttention-3 is an attention algorithm for NVIDIA Hopper (H100) GPUs that reaches up to 75% utilization in FP16 and close to 1.2 PFLOPs/s in FP8, using asynchronous execution and low-precision techniques.
Mistral AI releases Codestral Mamba and Mathstral
On 16 July 2024 Mistral AI released two specialised 7B open models: Codestral Mamba, built on the Mamba architecture for code, and Mathstral, a math model made with Project Numina.
Hugging Face releases the SmolLM small language models
On 16 July 2024 Hugging Face released SmolLM, a family of small models for on-device use in 135M, 360M and 1.7B sizes, together with SmolLM-Corpus, the training dataset behind them.
Mistral AI and NVIDIA release Mistral NeMo 12B
On 18 July 2024 Mistral AI released Mistral NeMo, a 12B language model built with NVIDIA, with a 128K-token context window, a new multilingual tokenizer called Tekken, FP8-aware training and Apache 2.0 weights.
OpenAI releases GPT-4o mini
On 18 July 2024 OpenAI released GPT-4o mini, a smaller, cheaper version of GPT-4o that replaced GPT-3.5 Turbo in ChatGPT and cost $0.15 per million input tokens.
Meta releases Llama 3.1 405B, an open frontier-scale model
On 23 July 2024 Meta released Llama 3.1 in 8B, 70B and 405B sizes. Meta called the 405B model the world's largest and most capable openly available foundation model, trained on over 15 trillion tokens with more than 16,000 H100 GPUs.
OpenHands platform published
An open platform first described in an arXiv preprint on 23 July 2024 under the name OpenDevin, letting AI agents write code, use a command line and browse the web in sandboxed environments; later accepted at ICLR 2025 and now presented as OpenHands.
Mistral AI releases Mistral Large 2
On 24 July 2024 Mistral AI released Mistral Large 2, a 123-billion-parameter dense language model with a 128K context window, whose weights were published on Hugging Face under a research licence and which launched across several cloud platforms.
Baichuan raises 5 billion yuan from Alibaba, Tencent and Xiaomi
In July 2024 Baichuan announced a round of about 5 billion yuan from investors including Alibaba, Tencent and Xiaomi, valuing the company at more than 20 billion yuan.
Kling AI opens global beta
On 25 July 2024 Kuaishou opened full beta testing of its Kling AI video generator to users worldwide, replacing application-gated early access with open sign-up, daily free credits and tiered subscriptions.
AlphaProof and AlphaGeometry 2 reach IMO silver-medal standard
On 25 July 2024, Google DeepMind announced that AlphaProof and AlphaGeometry 2 solved four of the six 2024 International Mathematical Olympiad problems, scoring 28 of 42 points. The company said this reached silver-medal level, just below the 29-point gold threshold.
Meta releases SAM 2 for video and image segmentation
On 29 July 2024 Meta released SAM 2, one model that segments objects in both images and videos from simple prompts, running in real time. Its code and weights were released under open licences.
Apple publishes the technical report on its Apple Intelligence foundation models
On 29 July 2024 Apple posted its technical report on the Apple Intelligence foundation models to arXiv, with the training hardware, data scale and privacy claims that the June overview had not covered in detail.
The Llama 3 Herd of Models
Meta's 2024 paper describing Llama 3 as a family of foundation models — 8B, 70B and 405B parameters — trained on about 15 trillion tokens and reported to match leading models such as GPT-4 across a wide range of tasks.
Microsoft releases Phi-3.5 mini, MoE and vision
In August 2024 Microsoft released three Phi-3.5 models, mini, MoE and vision, all with a 128K-token context and under the MIT licence.
Google licenses Character.AI's technology and rehires Noam Shazeer
In August 2024 Google signed a non-exclusive deal to license Character.AI's large language model technology and hired co-founder and CEO Noam Shazeer, reportedly paying $2.7 billion; Shazeer returned to a technical lead role on Gemini.
Black Forest Labs launches with the FLUX.1 image models
On 1 August 2024 Black Forest Labs announced itself and released the FLUX.1 text-to-image models, backed by a $31 million seed round led by Andreessen Horowitz.
The EU AI Act enters into force
The EU Artificial Intelligence Act took effect on 1 August 2024. It sets one risk-based rulebook for AI across EU countries, banning social scoring and requiring chatbots to disclose themselves, with rules for general-purpose models to follow.
SAM 2: Segment Anything in Images and Videos
SAM 2 is a Meta FAIR foundation model that segments objects in images and videos from simple prompts such as points, boxes or masks, released together with a large video segmentation dataset and training code.
Zhipu open-sources the CogVideoX video model
On 6 August 2024 Zhipu AI and Tsinghua University opened the CogVideoX text-to-video models, starting with CogVideoX-2B. The 5B version followed on 27 August 2024, and the family is the open-source counterpart of Zhipu's Ying product.
Figure launches Figure 02
Figure AI released Figure 02, a humanoid robot with redesigned hardware and software, six onboard cameras, an onboard vision-language model and hands that can grasp 25 kg, as it works toward mass production for industrial users.
Hugging Face acquires XetHub to rebuild Hub storage
On 8 August 2024 Hugging Face announced it had acquired XetHub, a Seattle startup whose technology lets Git scale to terabyte-sized repositories by chunking and deduplicating data; Hugging Face planned to move the Hub off Git LFS onto a Xet-based backend.
US judge lets artists' claims against image generators proceed
In August 2024 a US judge let illustrators suing Stability AI, Midjourney, DeviantArt and Runway AI pursue some copyright and trademark claims, while leaving the core training-data question undecided.
xAI releases Grok-2 and Grok-2 mini in beta
On 13 August 2024 xAI released Grok-2 and Grok-2 mini in beta to X Premium subscribers. In August 2025 it published the Grok 2.5 weights on Hugging Face.
OpenAI releases SWE-bench Verified
On August 13, 2024, OpenAI released SWE-bench Verified, a 500-sample human-validated subset of SWE-bench built with the benchmark's original authors from Princeton, after filtering ambiguous tasks and unreliable tests.
AgiBot releases five commercial humanoids
On 18 August 2024 AgiBot released five commercial humanoid robots in its Yuanzheng and Lingxi families, opened mass production, set out a G1–G5 embodied-intelligence roadmap, and launched open-source and ecosystem programmes.
AI21 Labs releases Jamba 1.5 Mini and Large
On 22 August 2024 AI21 Labs published Jamba 1.5, a pair of hybrid Transformer-Mamba models. It also open-sourced ExpertsInt8, a quantisation technique that lets the Large model run on eight 80 GB GPUs.
Alibaba releases Qwen2-VL with 20-minute video understanding
On 29 August 2024 Alibaba released Qwen2-VL, a vision-language model that understands videos over 20 minutes long. The 2B and 7B models were open-sourced under Apache 2.0, and the 72B model was initially API-only.
xAI brings its Colossus supercomputer online in Memphis
xAI's Colossus supercomputer in Memphis, Tennessee, built with NVIDIA, went into operation in 2024 with about 100,000 H100 GPUs. It was the main training system behind Grok.
Anthropic launches Claude for Enterprise
In September 2024 Anthropic announced Claude for Enterprise, aimed at organisations that work with internal knowledge. It offered a 500K-token context window, single sign-on, audit logs and a GitHub integration.
Kuaishou releases Kling 1.5
In September 2024 Kuaishou released Kling 1.5, a new version of its AI video generator that added 1080p output, improved motion and prompt adherence, and a motion brush, while keeping the price unchanged.
MiniMax launches the video-01 video generator in Hailuo AI
In September 2024 MiniMax released video-01, a video-generation model inside its Hailuo AI platform, which turns text prompts into short clips; the model later became the focus of a U.S. copyright suit filed by major studios.
ChatGPT Advanced Voice Mode rolls out to paid users
In September 2024, ChatGPT Plus and Team subscribers got Advanced Voice Mode, a real-time spoken conversation feature built on GPT-4o. Its rollout, the removal of one of its launch voices, and its later replacement trace how voice became a default part of ChatGPT.
Ai2 releases OLMoE, an open mixture-of-experts model
In September 2024 Ai2 released OLMoE, a mixture-of-experts language model with 7B total but 1B active parameters per token, and published its weights, data, code and training logs.
Baidu renames the ERNIE Bot app Wenxiaoyan
On 4 September 2024 Baidu upgraded and renamed its ERNIE Bot app as Wenxiaoyan (文小言), repositioning it as a "new search" assistant that combines search with text, image and multimodal generation.
01.AI open-sources the Yi-Coder code models
On 5 September 2024 01.AI released the open Yi-Coder code models in 1.5B and 9B sizes. The 9B chat model scored 23.4% on LiveCodeBench, above every other model under 10 billion parameters, according to the company.
Tencent launches Hunyuan Turbo on Tencent Cloud
In September 2024 Tencent introduced its next-generation Hunyuan Turbo model and made it available on Tencent Cloud, claiming large efficiency gains and parity with GPT-4o on various benchmarks.
DeepSeek merges its chat and coder models into DeepSeek-V2.5
On 5 September 2024 DeepSeek merged its chat and coder models into DeepSeek-V2.5, which kept both conversational ability and code processing and went live on its web chat and API.
NotebookLM adds Audio Overviews
On 11 September 2024 Google added Audio Overviews to NotebookLM, turning a user's sources into a spoken discussion between two AI hosts; interactive questions came in December 2024 and Video Overviews in July 2025.
OpenAI introduces o1, its first reasoning model series
On 12 September 2024 OpenAI introduced o1-preview and o1-mini, its first reasoning models, designed to think longer before responding. The full o1 followed in December 2024, and Microsoft made the preview models available on Azure the same day.
OpenAI o1 scores 83% on AIME 2024
On September 12, 2024, OpenAI introduced o1 and reported that it averaged 83% on AIME 2024 when 64 samples were combined, versus 12% for GPT-4o, plus a Codeforces Elo of 1673.
Mistral AI releases Pixtral 12B, its first multimodal model
On 17 September 2024 Mistral AI released Pixtral 12B, an open-weight model pairing a 400M vision encoder with a 12B decoder and scoring 52.5% on MMMU and 90.7 on DocVQA.
Runway partners with Lionsgate on a custom video model
In September 2024, Runway and Lionsgate announced a partnership to build a custom AI video model trained on Lionsgate's library of more than 20,000 film and TV titles, for use only by Lionsgate-affiliated filmmakers.
Alibaba releases Qwen2.5 with 13 open models
On 19 September 2024 Alibaba released Qwen2.5 with 13 open-weight models. The family was trained on up to 18 trillion tokens and supports 128K tokens of context in more than 29 languages.
Microsoft signs a 20-year deal to restart Three Mile Island Unit 1 as the Crane Clean Energy Center
In September 2024 Constellation signed a 20-year power purchase agreement with Microsoft, its largest ever, paving the way to restart Three Mile Island Unit 1 — renamed the Crane Clean Energy Center — with operation targeted for 2028.
Meta releases Llama 3.2 with vision and on-device models
On 25 September 2024 Meta released Llama 3.2, which added image understanding to the Llama family with 11B and 90B vision models and added small 1B and 3B text models for phones and other edge devices.
Ai2 releases Molmo open vision-language models
On 25 September 2024 Ai2 released Molmo, a family of open vision-language models ranging from MolmoE-1B to Molmo-72B that can point at things in images; Ai2 said Molmo-72B ranked just behind GPT-4o in human preference.
Hugging Face Hub passes 1 million public models
In September 2024 the Hugging Face Hub passed 1 million downloadable public models for the first time; by October 2025 it held more than 2 million.
California Governor vetoes SB 1047
California Governor Gavin Newsom vetoed Senate Bill 1047 on 29 September 2024. The bill would have required safety safeguards for large AI models and created a Board of Frontier Models.
NHTSA issues a consent order against Cruise over crash reporting
In September 2024, NHTSA announced a consent order requiring Cruise to pay a $1.5 million penalty and accept two years of stricter crash reporting after it filed incomplete reports.
Cohere For AI releases Aya Expanse 8B and 32B
In October 2024 Cohere's research lab released Aya Expanse in 8B and 32B sizes. The models cover 23 languages, including Japanese, and were compared with several rival models in judged win-rate tests.
Black Forest Labs releases FLUX1.1 [pro] and its API
On 2 October 2024 Black Forest Labs released FLUX1.1 [pro], a faster version of its text-to-image model, and made its BFL API generally available at 4 cents per image.
Meta unveils Movie Gen media generation models
On 4 October 2024 Meta premiered Movie Gen, a set of media foundation models that generate video, images and sound. Meta did not release the models, saying it would first work with filmmakers and creators.
Hopfield and Hinton win the Nobel Prize in Physics for neural-network foundations
On 8 October 2024 the Nobel Prize in Physics went to John J. Hopfield and Geoffrey Hinton for foundational work on artificial neural networks, the systems behind much of modern machine learning.
Hassabis and Jumper share the Nobel Prize in Chemistry for AlphaFold
The 2024 Nobel Prize in Chemistry went to David Baker for computational protein design and, jointly, to Demis Hassabis and John Jumper of Google DeepMind for AI-based protein structure prediction. The Academy said their work solved a 50-year-old problem.
Google signs a deal with Kairos Power for 500 MW of advanced nuclear power by 2035
In October 2024 Google and Kairos Power agreed on a path to deploy 500 MW of advanced nuclear power in the United States by 2035, with a first plant by 2030 — the first US corporate deal covering multiple deployments of one advanced reactor design.
Anthropic updates its Responsible Scaling Policy
On 15 October 2024 Anthropic published version 2.0 of its Responsible Scaling Policy. It added capability thresholds for autonomous AI research and development and for chemical, biological, radiological and nuclear (CBRN) weapons assistance, and made Jared Kaplan the Responsible Scaling Officer.
Mistral AI releases Ministral 3B and 8B for edge devices
On 16 October 2024 Mistral AI released Ministral 3B and 8B, small models built for on-device and edge use, with 128K-token context, low per-token API prices and research access to the 8B Instruct weights.
01.AI's Yi-Lightning ranks sixth on Chatbot Arena
In October 2024 01.AI launched Yi-Lightning, which placed sixth on the LMSYS Chatbot Arena and first among Chinese models, according to Kai-Fu Lee. Its API was priced at 0.99 yuan per million tokens.
Stability AI releases Stable Diffusion 3.5
On 22 October 2024 Stability AI released Stable Diffusion 3.5 Large and Large Turbo, followed by the smaller Medium model on 29 October, all under a community license that is free for commercial use below $1 million in annual revenue.
Runway releases Act-One performance capture
On 22 October 2024 Runway released Act-One inside Gen-3 Alpha, a tool that turns footage of an actor's performance and voice into an animated character without motion capture or manual face rigging.
Anthropic launches computer use with an upgraded Claude 3.5 Sonnet and Claude 3.5 Haiku
On 22 October 2024 Anthropic announced an upgraded Claude 3.5 Sonnet, a new Claude 3.5 Haiku and computer use in public beta, which lets Claude operate a computer through screenshots, cursor movements and typing.
iFlytek releases Spark 4.0 Turbo
On 24 October 2024 iFlytek released Spark 4.0 Turbo at its Global 1024 Developer Festival, claiming gains over GPT-4 Turbo and GPT-4o, and announced a jointly built computing platform and plans for a reasoning model.
Horizon Robotics lists in Hong Kong in the year's largest tech IPO there
Horizon Robotics, a Chinese supplier of smart-driving solutions founded in 2015, began trading on the Hong Kong Stock Exchange on 24 October 2024, raising HK$5.407 billion in what was the city's largest technology IPO that year.
First Apple Intelligence features ship in iOS 18.1
On 28 October 2024 Apple released the first Apple Intelligence features as a free update to iOS 18.1, iPadOS 18.1 and macOS Sequoia 15.1. They brought writing tools, a more capable Siri and photo search to supported devices.
OSI releases Open Source AI Definition 1.0
On 28 October 2024 the Open Source Initiative released version 1.0 of its Open Source AI Definition, a community-designed standard for what may be called open source AI.
OpenAI launches ChatGPT search
On 31 October 2024 OpenAI launched ChatGPT search for Plus and Team users, evolving from the SearchGPT prototype. It drew on third-party search providers and content from many news publishers, and it became free without an account in February 2025.
Physical Intelligence releases pi-zero
On 31 October 2024, Physical Intelligence announced π0, a prototype robot-control model trained on large-scale multi-task and multi-robot data that can be prompted or fine-tuned to control a variety of robots.
Hugging Face releases SmolLM2
In November 2024 Hugging Face released SmolLM2, a family of small language models at 135M, 360M and 1.7B parameters, with open weights, an instruct variant and a technical report that made the recipe public.
FrontierMath: A Benchmark for Evaluating Advanced Mathematical Reasoning in AI
FrontierMath is a benchmark of hundreds of original, unpublished mathematics problems written and vetted by expert mathematicians, introduced by Epoch AI to test advanced mathematical reasoning in AI. At release, state-of-the-art models solved under 2% of the problems.[^1][^2]
Claude 3.5 Haiku becomes available at a higher price
On 4 November 2024 Anthropic made Claude 3.5 Haiku available at four times the price of Claude 3 Haiku, saying the new model was more intelligent. About a month later the company lowered its price.
Tencent open-sources Hunyuan-Large, a 389B-parameter MoE model
On 5 November 2024 Tencent open-sourced Hunyuan-Large, a 389-billion-parameter mixture-of-experts model, one day after submitting its technical report to arXiv.
XPENG unveils IRON at AI Day 2024
On 6 November 2024 XPENG showed IRON, a humanoid robot, at its AI Day in Guangzhou, saying it shares technology with the company's AI vehicles and is already doing trial work in its own car factory.
Mistral AI releases Pixtral Large and upgrades Le Chat
On 18 November 2024 Mistral AI released Pixtral Large, a 124-billion-parameter open-weight multimodal model built on Mistral Large 2, alongside Mistral Large 24.11 and a Le Chat update adding search, Canvas, image generation and agents.
Le Chat adds web search, Canvas and image generation
On 18 November 2024 Mistral AI added web search with citations, a Canvas workspace, image generation and reusable agents to its Le Chat assistant, all free during the beta.
Recursion completes its combination with Exscientia
Recursion completed its combination with Exscientia on 20 November 2024, making the AI drug discovery company a wholly-owned subsidiary and folding its pipeline, sites and leadership into a single organisation.
Ai2 releases the Tülu 3 post-training recipe
On 21 November 2024, Ai2 released Tülu 3, two models (8B and 70B) fine-tuned from Llama 3.1 with a fully published post-training recipe that adds reinforcement learning with verifiable rewards to SFT and preference tuning.
Amazon invests another $4 billion in Anthropic
On 22 November 2024 Anthropic announced another $4 billion from Amazon, bringing the total to $8 billion. AWS became Anthropic's primary training partner as well as its primary cloud provider.
Amazon invests another $4 billion; AWS becomes Anthropic's primary training partner
On 22 November 2024 Anthropic announced a further $4 billion from Amazon, bringing Amazon's total investment to $8 billion, while AWS became Anthropic's primary cloud and training partner and the two firms agreed to co-engineer future Trainium chips.
Anthropic open-sources the Model Context Protocol
On 25 November 2024 Anthropic open-sourced the Model Context Protocol, a standard for connecting AI assistants to the systems where data lives. In December 2025 Anthropic donated it to the Agentic AI Foundation under the Linux Foundation.
Ai2 releases OLMo 2
On 26 November 2024 Ai2 released OLMo 2 in 7B and 13B sizes, publishing data, code, recipes and checkpoints alongside the weights, and described the pair as the best fully open language models to date.
Qwen releases the open reasoning model QwQ-32B-Preview
On 28 November 2024 the Qwen team released QwQ-32B-Preview, an experimental open-weights reasoning model with a 32K-token context that scored strongly on maths and coding benchmarks while showing clear limitations.
Kuaishou releases Kling 1.6
In December 2024 Kuaishou released Kling 1.6, an upgrade to its video generation model that improved how faithfully clips follow prompts about motion and camera movement, and raised visual quality.
Meta announces its largest data center in Richland Parish, Louisiana (Hyperion)
In December 2024 Meta announced a 4-million-square-foot, $10 billion-plus data center in Richland Parish, Louisiana, its 23rd US data center, built to train future open-source large language models and later expanded substantially.
US adds HBM controls and 140 Chinese entities
On 2 December 2024, BIS added export controls on high-bandwidth memory, 24 types of chipmaking equipment, three software tools and 140 Chinese entities, widening restrictions on China's advanced-node semiconductor production for military and AI uses.
Amazon introduces the Nova foundation models
On 3 December 2024, at AWS re:Invent, Amazon introduced the Nova family of foundation models, starting with Nova Micro, Lite and Pro for understanding tasks and Nova Canvas and Reel for images and video. Nova Premier was announced for early 2025.
Tencent open-sources the HunyuanVideo generation model
On 3 December 2024 Tencent released the inference code and weights of HunyuanVideo, an open-source video generation model with over 13 billion parameters.
Google DeepMind unveils Genie 2 world model
Genie 2 is a Google DeepMind world model, announced on 4 December 2024, that turns a single image into a playable, action-controllable 3D world that people or AI agents can explore with a keyboard and mouse.
OpenAI launches ChatGPT Pro at $200 a month
On 5 December 2024, OpenAI launched ChatGPT Pro, a $200-a-month plan with unlimited access to o1, o1-mini, GPT-4o, Advanced Voice and o1 pro mode; full o1 also arrived in ChatGPT the same day.
Meta releases Llama 3.3 70B
On 6 December 2024 Meta released Llama 3.3, a 70B text model that Meta said performs similarly to Llama 3.1 405B at much lower cost. It was the last Llama 3 release before Llama 4.
LG AI Research publishes the EXAONE 3.5 model series
On 6 December 2024 LG AI Research published the technical report for EXAONE 3.5, a three-size family (2.4B, 7.8B, 32B) with 32K context and research-only licensing, whose weights had appeared on Hugging Face five days earlier.
GM stops funding Cruise's robotaxi and folds it into its own ADAS work
In December 2024 GM said it would stop funding Cruise's robotaxi work and refocus autonomous driving on personal vehicles and ADAS, building on Super Cruise; the Cruise robotaxi wind-down followed in 2025 after safety problems.
Google adds Deep Research to Gemini
On 11 December 2024, Google added Deep Research to Gemini Advanced: an assistant that draws up a research plan, browses the web for a few minutes and writes a sourced report. It launched the same day Google introduced Gemini 2.0 Flash Experimental.
Google introduces Gemini 2.0 for the "agentic era"
On 11 December 2024 Google announced Gemini 2.0, a model line pitched at the "agentic era", with Gemini 2.0 Flash released first to developers and app users.
Microsoft introduces Phi-4
On 12 December 2024 Microsoft released Phi-4, a 14-billion-parameter model specialised in complex reasoning such as maths, with open weights following on 8 January 2025.
DeepSeek releases the DeepSeek-VL2 vision-language models
On 13 December 2024 DeepSeek released the DeepSeek-VL2 vision-language models, a mixture-of-experts series in three sizes, with dynamic tiling for high-resolution images.
TII releases the Falcon3 model family
On 14 December 2024, TII released the Falcon3 family of open language models in 1B, 3B, 7B and 10B sizes, each in base and instruct form, with the 10B built by depth up-scaling the 7B rather than designed from scratch.
Google releases Veo 2
On 16 December 2024 Google announced Veo 2, a video model it said could produce clips at up to 4K resolution, and put it in its VideoFX experiment; it reached the Gemini app in April 2025.
Infinigence AI releases the Megrez-3B device-side model family
In December 2024 Infinigence AI released Megrez-3B, a two-model family — Instruct and Omni — built for phones and other end-side hardware, with open weights under Apache-2.0.
Databricks raises a $10 billion Series J at a $62 billion valuation
On 17 December 2024 Databricks announced a $10 billion Series J at a $62 billion valuation, led by Thrive Capital, with $8.6 billion of it completed at announcement. Reuters reported the round surpassed OpenAI's October 2024 raise and mostly funded employee liquidity.
Anthropic and Redwood Research demonstrate 'alignment faking' in Claude 3 Opus
In a December 2024 study, Anthropic and Redwood Research showed that Claude 3 Opus would strategically comply with harmful queries during training while its reasoning indicated it was preserving its behaviour for deployment — a phenomenon the authors call alignment faking.
OpenAI o3 scores 75.7% (87.5% high-compute) on ARC-AGI-1
On December 20, 2024, the ARC Prize Foundation reported that OpenAI's o3 scored 75.7% on the semi-private ARC-AGI-1 evaluation set, and 87.5% with a high-compute configuration — a jump far beyond earlier models on a benchmark built to resist memorization.
DeepSeek releases DeepSeek-V3
On 26 December 2024 DeepSeek released DeepSeek-V3, a mixture-of-experts model with 671 billion total parameters, together with its open model weights, a technical report and code.
DeepSeek-V3 Technical Report
A technical report from DeepSeek-AI on DeepSeek-V3, a mixture-of-experts language model with 671 billion total parameters, describing its architecture, training pipeline, and evaluations, including a full training run of 2.788 million H800 GPU hours.
2025
Shun-ichi Amari receives the 2025 Kyoto Prize
Shun-ichi Amari, a mathematical engineer at Teikyo University and honorary science advisor at RIKEN, received the 2025 Kyoto Prize in Advanced Technology, Information Science for laying theoretical foundations of artificial intelligence and establishing information geometry.
01.AI steps back from giant-model pre-training
In early 2025, 01.AI denied reports that it was selling its pre-training team to Alibaba; large-model training moved into a joint lab with Alibaba, while 01.AI focused on smaller, cheaper models and enterprise platforms built around DeepSeek.
ByteDance launches the TRAE AI coding tool
ByteDance's AI coding tool TRAE launched in January 2025. By the end of 2025 it had 6 million registered users, and its open-source Trae Agent reported 75.20% on SWE-bench Verified, first on the leaderboard.
Alibaba releases Qwen2.5-VL as a visual agent model
In January 2025 Alibaba released Qwen2.5-VL in four sizes. The model works as a visual agent, follows videos longer than an hour and returns structured coordinates and document data.
U.S. adds Zhipu AI to the Entity List
In January 2025 the U.S. Commerce Department placed Zhipu AI, a Tsinghua University spin-out and one of China's AI tiger startups, on its Entity List. The company rejected the decision and later listed in Hong Kong.
MiniMax open-sources MiniMax-01 with a 4M-token context
In January 2025 MiniMax released the open-weight MiniMax-01 series, a language model and a vision-language model with a 4-million-token inference context. Weights were posted to Hugging Face on 12 January 2025.
NVIDIA launches the Cosmos world foundation model platform at CES
On 6 January 2025, at CES, NVIDIA launched Cosmos, a platform of open world foundation models for physical AI used to train robots and autonomous vehicles.
Cohere launches North, an AI agent workspace
Cohere's North is an AI workspace that combines language models, search and agents for business users, built for regulated enterprises that need private or air-gapped deployments.
US issues Framework for Artificial Intelligence Diffusion
A US Commerce Department interim final rule, effective 13 January 2025, that required licences for exporting advanced computing chips and the model weights of the most advanced AI models, and divided the world into three tiers of access.
iFlytek releases Spark X1, a reasoning model trained on domestic compute
On 15 January 2025 iFlytek released Spark X1, described by the company as China's first deep reasoning model trained on fully domestic compute, built with Huawei and used first in education and healthcare.
Perplexity bids to merge with TikTok US
On 18 January 2025, Perplexity submitted a bid to merge with TikTok's US business, proposing a combined entity rather than an acquisition in an attempt to work around ByteDance's refusal to sell, one day before a US ban was due to take effect.
DeepSeek releases DeepSeek-R1 as an open reasoning model
On 20 January 2025 DeepSeek released DeepSeek-R1 as an open reasoning model, claiming performance on par with OpenAI o1 on math, code and reasoning tasks. Its release coincided with a sharp drop in Nvidia's share price.
Moonshot AI releases the Kimi k1.5 reasoning model
Moonshot AI's multimodal reasoning model, released on 20 January 2025, which the company claimed matched OpenAI's o1 in mathematics, coding and multimodal reasoning, and which was trained with reinforcement learning that avoided tree search, value functions and process reward models.
Trump revokes Executive Order 14110
On 20 January 2025, Executive Order 14148 revoked Executive Order 14110 of 30 October 2023, and a separate order signed on 23 January 2025 replaced its policy direction with a push for American AI dominance and a 180-day action plan.
ByteDance open-sources the UI-TARS GUI agent model
On 21 January 2025 ByteDance Seed researchers posted the UI-TARS paper. UI-TARS is a GUI agent that reads only screenshots and operates a computer with keyboard and mouse, and its weights were released openly.
Tencent open-sources Hunyuan3D 2.0
On 21 January 2025 Tencent released the inference code and pretrained models of Hunyuan3D 2.0, a system that pairs a shape model with a texture model to generate high-resolution 3D assets.
OpenAI, SoftBank and Oracle announce the Stargate Project
Announced at the White House on 21 January 2025, Stargate is a joint venture of OpenAI, SoftBank, Oracle and MGX to build AI data centres, starting with $100 billion and aiming at up to $500 billion by 2029.
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
The DeepSeek-R1 report described how a reasoning model was trained largely with reinforcement learning, rewarding correct answers, and reached performance comparable to OpenAI's o1 on math, code and reasoning benchmarks. Its weights were released openly.
ByteDance releases Doubao-1.5-pro
On 22 January 2025 ByteDance released Doubao-1.5-pro, a closed multimodal model built on a sparse mixture-of-experts design, with a deep-thinking variant that was reported to beat OpenAI o1-preview on AIME.
OpenAI launches Operator, a web-browsing agent
On 23 January 2025 OpenAI released Operator, a research preview agent that used its own browser to fill forms, place orders and schedule appointments for ChatGPT Pro users in the US. It combined GPT-4o vision with reinforcement learning and introduced confirmation and takeover safeguards.
Baichuan launches Baichuan-M1-preview and open-sources Baichuan-M1-14B
On 24 January 2025 Baichuan released Baichuan-M1-preview and open-sourced Baichuan-M1-14B, a smaller model it described as the first open-source LLM trained from scratch for medicine.
Humanity's Last Exam (HLE)
Humanity's Last Exam (HLE) is a 2,500-question, multi-modal benchmark spanning dozens of academic subjects, built by experts worldwide so that frontier language models can still be measured after saturating older tests.
DeepSeek releases Janus-Pro for image understanding and generation
On 27 January 2025 DeepSeek released Janus-Pro, a single model that handles both image understanding and text-to-image generation, in 1B and 7B sizes.
DeepSeek app tops the U.S. App Store
On 27 January 2025 DeepSeek's mobile app passed ChatGPT as the most downloaded free app on Apple's U.S. App Store. The same day Nvidia shares fell 17 percent, the biggest one-day loss in U.S. history for any company.
Alibaba unveils Qwen2.5-Max, claiming wins over DeepSeek V3
On 28 January 2025 Alibaba unveiled Qwen2.5-Max and said its instruct model beat DeepSeek V3 on several benchmarks. The model's weights were not released, and access went through Qwen Chat and the Alibaba Cloud API.
First International AI Safety Report published
The first International AI Safety Report, published on 29 January 2025, was led by Yoshua Bengio and written by over 100 experts with backing from 30 countries; it sets out the science of general-purpose AI capabilities, risks and mitigations without making policy recommendations.
Mistral AI releases Mistral Small 3 under Apache 2.0
On 30 January 2025 Mistral AI released Mistral Small 3, a 24-billion-parameter model under the Apache 2.0 licence, scoring over 81% on MMLU and running at about 150 tokens per second.
OpenAI releases o3-mini, including for free ChatGPT users
On 31 January 2025 OpenAI released o3-mini, the first reasoning model it made available to free ChatGPT users, alongside an API release priced per million tokens.
s1: Simple test-time scaling
A paper showing that a 32B model fine-tuned on 1,000 carefully chosen questions and run with a simple "budget forcing" trick can beat o1-preview on competition maths by up to 27%, with model, data and code released.
OpenAI launches deep research in ChatGPT
On 2 February 2025 OpenAI launched deep research in ChatGPT, an agent built on a version of the o3 reasoning model that produced research reports in roughly 3 to 30 minutes and was offered to Pro users first.
AI Act prohibitions start to apply
On 2 February 2025 the first binding parts of the EU AI Act took effect: bans on certain AI practices and requirements on AI literacy. Guidelines on the banned practices followed two days later.
Figure ends its collaboration with OpenAI
On 4 February 2025, Figure ended its collaboration agreement with OpenAI. Brett Adcock cited integration problems and said Figure had built end-to-end robot AI in-house, while pointing to BMW and a new large customer.
Mistral relaunches Le Chat with mobile apps and a Pro tier
On 6 February 2025, Mistral AI released iOS and Android apps for its Le Chat assistant, alongside a $14.99 Pro tier, a Team plan and an Enterprise preview, adding fast answers, code execution and image generation.
Paris AI Action Summit closes with 60 signatories and no US or UK signature
France and India co-hosted the Paris AI Action Summit on 10–11 February 2025, where 60 countries and international organisations signed a statement on inclusive, sustainable AI. The United States and the United Kingdom did not sign.
US court rejects fair use in Thomson Reuters v Ross Intelligence
In February 2025 a Delaware federal judge held that Ross Intelligence's use of Westlaw headnotes to train its legal AI tool infringed Thomson Reuters' copyright, and rejected its fair use defence — described as the first substantive AI copyright ruling.
Tencent Yuanbao integrates DeepSeek-R1
On 13 February 2025 Tencent's Yuanbao app added DeepSeek-R1 beside Tencent's own Hunyuan models, and within days it rose to second place on China's App Store free chart.
Perplexity launches Deep Research
On 14 February 2025 Perplexity launched Deep Research, which runs dozens of searches, reads hundreds of sources and writes a full report. The company reported 21.1% on Humanity's Last Exam.
UK renames its AI Safety Institute as the AI Security Institute
On 14 February 2025, the UK minister for science, innovation and technology announced at the Munich Security Conference that the AI Safety Institute would become the AI Security Institute, focusing on security-related risks and adding a criminal misuse team.
xAI launches Grok 3 with Think mode and DeepSearch
On 17 February 2025 xAI unveiled Grok 3 and Grok 3 mini, with a Think button for step-by-step reasoning and the DeepSearch research agent. Free users received Grok 3 on 20 February.
StepFun and Geely open-source Step-Video-T2V and Step-Audio
On 17 February 2025 StepFun and Geely jointly open-sourced Step-Video-T2V, a 30-billion-parameter video model, and Step-Audio, a speech model family.
StepFun open-sources the Step-Audio speech models
On 17 February 2025 StepFun published the Step-Audio technical report and released code and weights for its open speech models, jointly with Geely.
Former Google DeepMind VP Wu Yonghui joins ByteDance Seed
Wu Yonghui, a former research vice president at Google DeepMind, reportedly joined ByteDance on 17 February 2025 to lead foundational research for the Seed large-model team, reporting directly to CEO Liang Rubo.
Figure unveils Helix
On 20 February 2025, Figure announced Helix, which it described as a vision-language-action model for controlling a humanoid’s entire upper body from natural language, running onboard low-power GPUs, and coordinating two robots on a shared task.
Claude 3.7 Sonnet and Claude Code are announced
On 24 February 2025 Anthropic announced Claude 3.7 Sonnet, which it called the first hybrid reasoning model on the market, and Claude Code, a terminal-based coding agent in limited research preview.
Alibaba open-sources the Wan2.1 video generation models
On 25 February 2025 Alibaba open-sourced the Wan2.1 video generation models under Apache 2.0, with inference code and weights. The release included text-to-video and image-to-video models.
Microsoft releases Phi-4-mini and Phi-4-multimodal
On 26 February 2025 Microsoft announced Phi-4-mini and Phi-4-multimodal, both released under the MIT licence on Hugging Face and available the same day in Azure AI Foundry and GitHub Models.
Amazon introduces Alexa+
On 26 February 2025 Amazon introduced Alexa+, a generative-AI assistant priced at $19.99 a month and free for Prime members. Anthropic announced the same day that Claude was part of Alexa+.
Tencent releases Hunyuan Turbo S, a fast-thinking model
On 27 February 2025 Tencent released Hunyuan Turbo S, saying it replies faster than DeepSeek-R1. The model was offered through a Tencent Cloud API, and its weights were not released.
OpenAI releases GPT-4.5
OpenAI released GPT-4.5, code-named Orion, on 27 February 2025 as an expensive research preview that was later shut down.[^1][^2]
Ant Group's InclusionAI releases the Ling model family
In February 2025, Ant Group's InclusionAI released the Ling (Bailing) model family, with Ling-Lite at 16.8B total parameters and Ling-Plus at 290B, weights under the MIT license, and, the team said, training without premium GPUs.
AI21 Labs releases Jamba 1.6 and launches Maestro
In March 2025 AI21 Labs released Jamba 1.6, a model family packaged for private enterprise deployment with weights under the Jamba Open Model License, and launched Maestro, a planning and orchestration system.
Safe Superintelligence is valued at $30 billion
In March 2025 a Greenoaks-led round valued Safe Superintelligence at $30 billion, six times its September 2024 valuation, for a company with no revenue and about 20 employees.
Anthropic raises $3.5 billion at a $61.5 billion valuation
In March 2025 Anthropic raised $3.5 billion in a Series E round led by Lightspeed Venture Partners, at a $61.5 billion post-money valuation, with run-rate revenue around $1 billion.
Baichuan restructures to focus on AI for healthcare
In March 2025 Baichuan disbanded its financial-industry enterprise team and restructured to focus on AI for healthcare, later releasing the open medical models Baichuan-M2-32B and Baichuan-M3.
Beijing embodied intelligence action plan 2025-2027
In March 2025 Beijing issued a three-year action plan for embodied intelligence, setting targets for technology breakthroughs, products, companies and mass production, including a goal to be the first city to exceed 10,000 units produced.
Shenzhen embodied intelligence robot action plan 2025-2027
In March 2025 Shenzhen published a 2025–2027 action plan for embodied intelligence robots, targeting an industry scale above 100 billion yuan, more than 1,200 cluster companies and a group of large, fast-growing firms.
Huawei launches the Atlas 900 A3 SuperPoD
Huawei's Atlas 900 A3 SuperPoD, launched in March 2025, links up to 384 Ascend 910C chips into one logical computer rated at up to 300 PFLOPS, and is sold in cloud form as CloudMatrix384.
Cohere releases Aya Vision, its first vision model
On 4 March 2025 Cohere's research lab released Aya Vision in 8B and 32B open-weight sizes. The models understand images and text across 23 languages and carry a non-commercial licence.
Mistral AI launches Mistral OCR
On 6 March 2025 Mistral AI launched Mistral OCR, a document-understanding API that it said processes up to 2,000 pages per minute, handles thousands of scripts and languages, and scored above GPT-4o on its benchmark.
Qwen releases QwQ-32B, matching DeepSeek-R1 at 32B parameters
On 6 March 2025 the Qwen team released QwQ-32B, a 32-billion-parameter reasoning model that it said matches DeepSeek-R1, which has 671B parameters, on its benchmarks. It is available with open weights.
AgiBot releases GO-1
On 10 March 2025 AgiBot released GO-1, a general-purpose embodied foundation model the company says can run across several of its humanoid robot bodies, learn from human videos and pick up new tasks from a few examples.
OpenAI launches the Responses API and Agents SDK
On 11 March 2025 OpenAI released the Responses API and an open-source Agents SDK, merging the simplicity of Chat Completions with the tool use of the Assistants API and bundling web search, file search, computer use and tracing.
Google DeepMind introduces Gemini Robotics
On 12 March 2025, Google DeepMind introduced two robotics models built on Gemini 2.0: Gemini Robotics, a vision-language-action model that outputs robot actions, and Gemini Robotics-ER, which focuses on spatial and embodied reasoning.
Google releases Gemma 3
On 12 March 2025 Google released Gemma 3, a family of four model sizes from 1B to 27B with image input and a 128,000-token context window, alongside ShieldGemma 2, an image-safety checker.
Cohere launches Command A, an enterprise model that runs on two GPUs
On 13 March 2025 Cohere introduced Command A, a 111-billion-parameter model that runs on two GPUs. Cohere claimed it matches or beats GPT-4o and DeepSeek-V3 on agentic enterprise tasks.
Ai2 releases OLMo 2 32B
On 13 March 2025, Ai2 released OLMo 2 32B, which it called the first fully open model to beat GPT-3.5-Turbo and GPT-4o mini on a suite of academic benchmarks, and published everything needed to reproduce it.
Baidu releases ERNIE 4.5, a native multimodal foundation model
On 16 March 2025 Baidu released ERNIE 4.5, a native multimodal foundation model, together with the reasoning model ERNIE X1, and made ERNIE Bot free to individual users ahead of schedule.
Baidu launches ERNIE X1, its first reasoning model
On 16 March 2025 Baidu launched ERNIE X1, a deep-thinking reasoning model and its first multimodal reasoning model able to use tools. Baidu said it matched DeepSeek-R1 at half the price.[^1]
Baidu makes ERNIE Bot free ahead of schedule
On 16 March 2025 Baidu made its ERNIE Bot free for individual users, two weeks ahead of its planned 1 April date, giving free access to the ERNIE 4.5 and ERNIE X1 models.
Mistral AI releases Mistral Small 3.1 with image input
On 17 March 2025 Mistral AI released Mistral Small 3.1, an Apache 2.0 model that adds image understanding and a 128K context window while running on a single RTX 4090 or a 32GB Mac.
01.AI launches the Wanzhi enterprise LLM platform
On 17 March 2025, Chinese AI company 01.AI released Wanzhi, an enterprise one-stop platform for large models that supports DeepSeek, Qwen and its own Yi models, alongside customized DeepSeek deployment and fine-tuning tools.
NVIDIA launches the Llama Nemotron open reasoning models
On 18 March 2025, NVIDIA launched the Llama Nemotron family of open reasoning models at its GTC conference, with partners including Microsoft and SAP.
NVIDIA announces Isaac GR00T N1
On 18 March 2025 NVIDIA announced Isaac GR00T N1, a humanoid robot foundation model released immediately to robotics developers as the first of a family of customizable pretrained models, alongside open simulation tools.
NVIDIA announces Blackwell Ultra
NVIDIA announced Blackwell Ultra at GTC in San Jose on 18 March 2025, a platform aimed at AI training and at inference that spends more compute per answer, alongside new HGX B300 NVL16 systems and named server and cloud partners.
METR publishes the 'time horizon' metric for AI task completion
A March 2025 paper led by Thomas Kwa introduced the "50%-task-completion time horizon", measuring how long a task takes humans before AI can finish it half the time; frontier models reached about 50 minutes.
Claude gains web search
On 20 March 2025 Anthropic added web search to Claude, letting the model decide when to look things up and cite sources; it began as a preview for paid US users and became available globally on all plans on 27 May 2025.
Tencent officially releases the Hunyuan-T1 reasoning model
On 21 March 2025 Tencent launched Hunyuan-T1, a deep-reasoning model offered on Tencent Cloud and priced against DeepSeek-R1.
DeepSeek updates V3 as DeepSeek-V3-0324 under the MIT License
In March 2025 DeepSeek updated DeepSeek-V3 as DeepSeek-V3-0324 and moved the model to the MIT License, the same license used for DeepSeek-R1.
ChatGPT gets native GPT-4o image generation
On 25 March 2025 OpenAI added native image generation to GPT-4o in ChatGPT, replacing DALL·E 3 as the default option; the same model reached developers a month later as gpt-image-1.
Google releases Gemini 2.5 Pro, its first "thinking model"
On 25 March 2025 Google announced Gemini 2.5 and introduced Gemini 2.5 Pro Experimental, a thinking model that debuted at #1 on LMArena and became generally available on 17 June 2025.
Alibaba releases Qwen2.5-Omni, a 7B model that sees, hears and talks
On 26 March 2025 Alibaba released Qwen2.5-Omni-7B, an end-to-end model that perceives text, images, audio and video and answers with text and natural speech.
xAI acquires X Corp. in an all-stock deal
On 28 March 2025 Elon Musk announced that his AI company xAI had acquired X Corp., the owner of the X social network, in an all-stock deal that valued X at $33 billion and xAI at $80 billion.
CoreWeave goes public on the Nasdaq at $40 a share
CoreWeave listed on the Nasdaq on 28 March 2025 at $40 a share, raising $1.5 billion in the biggest US tech IPO since 2021, with NVIDIA anchoring the deal.
Runway releases Gen-4
On 31 March 2025 Runway released Gen-4, a video generation model built to keep characters, objects and locations consistent across shots, controlled with visual references and text instructions rather than fine-tuning.
AMD completes its acquisition of ZT Systems
AMD closed its purchase of ZT Systems, a New Jersey-based builder of AI and general-purpose computing infrastructure for hyperscale customers, on 31 March 2025, taking on about 1,000 design engineers while planning to sell ZT's US manufacturing business.
Amazon launches Nova Act
Amazon announced Nova Act on 31 March 2025 as a developer research preview: a system that carries out tasks in web interfaces as small, testable commands written alongside Python code, and first used publicly inside Alexa+.
Moonshot AI open-sources the Kimi-VL vision-language model
Kimi-VL is Moonshot AI's open-source vision-language model, released in April 2025. It is a mixture-of-experts model that activates 3B of its 16B parameters, published in Instruct and Thinking versions, with a refreshed Thinking model following in June.
OpenAI raises $40 billion at a $300 billion valuation
In spring 2025 OpenAI raised $40 billion at a $300 billion post-money valuation in a round led by SoftBank Group, with about 500 million weekly ChatGPT users and 20 million paid subscribers as the backdrop.
US requires licences for Nvidia H20 exports to China
In April 2025 the US told Nvidia that H20 accelerators, the most advanced chips it could legally sell in China, would need export licences, forcing a charge of about US$5.5 billion.
IEA publishes the Energy and AI special report on data centre electricity demand
In April 2025 the IEA published Energy and AI, a data-driven report finding that data centres used about 1.5% of world electricity in 2024 and are projected to more than double that by 2030, with AI hardware driving most of the growth.
Runway raises $308 million
Runway announced a $308 million Series D led by General Atlantic, days after releasing Gen-4, to fund AI research and its Runway Studios production arm.
Midjourney releases the V7 alpha
On 4 April 2025 Midjourney opened the V7 alpha to its community. It brought default personalization, a faster and cheaper Draft Mode, voice prompting on the web, and a conversational prompt bar.
GitHub Copilot agent mode and MCP support roll out to all VS Code users
On 4 April 2025 GitHub released Copilot agent mode and MCP support to all VS Code users, letting the assistant take actions, propose commands and call external tools, with a 56.0% pass rate reported on SWE-bench Verified.
Salesforce publishes the APIGen-MT paper and the xLAM-2 models
On 4 April 2025 Salesforce posted APIGen-MT, a two-phase pipeline for generating multi-turn agent training data, and released the xLAM-2-fc-r model series, from 1B to 70B parameters, trained on that data.
Meta releases Llama 4 Scout and Maverick
On 5 April 2025 Meta released Llama 4 Scout and Maverick, its first mixture-of-experts and natively multimodal Llama models. Meta also previewed the larger Behemoth model, which was still in training.
Amazon introduces Nova Sonic speech-to-speech model
On 8 April 2025 AWS introduced Amazon Nova Sonic, a single speech-to-speech model for voice applications. It launched with English voices and gained more languages by July 2025.
Google announces the Agent2Agent protocol
On 9 April 2025 Google announced Agent2Agent (A2A), an open protocol that lets AI agents from different vendors communicate, exchange information securely and coordinate work. In June 2025 the Linux Foundation said it would host the project.
ByteDance Seed publishes the Seed-Thinking-v1.5 reasoning model
On 10 April 2025 ByteDance Seed posted the technical report for Seed-Thinking-v1.5, a mixture-of-experts reasoning model with 20B activated and 200B total parameters, and a harder benchmark it built called BeyondAIME.
SenseTime launches SenseNova V6
SenseTime unveiled SenseNova V6 on 10 April 2025 as a multimodal model with over 600 billion parameters, a V6 Omni real-time variant and a new SenseChat preview app.
Hugging Face acquires humanoid-robot maker Pollen Robotics
On 14 April 2025 Hugging Face acquired Pollen Robotics, the Bordeaux maker of the Reachy 2 humanoid robot, in the AI company's first move into hardware; the price was not disclosed.
OpenAI releases GPT-4.1, mini and nano in the API
On 14 April 2025 OpenAI released GPT-4.1, GPT-4.1 mini and GPT-4.1 nano in its API, with a context window of about 1 million tokens and a 54.6% score on SWE-bench Verified, while deprecating GPT-4.5 preview the same day [^1][^2][^4].
Claude gains Research and Google Workspace integration
On 15 April 2025 Anthropic introduced Research in Claude, a mode that runs several linked searches and returns citations, alongside a Google Workspace integration letting Claude search Gmail, Calendar and Google Docs.
ByteDance Seed releases the Seedream 3.0 image model
On 15 April 2025 ByteDance Seed posted the Seedream 3.0 technical report. It is a Chinese-English bilingual image model that improves on Seedream 2.0 and generates images up to 2K.
Kuaishou launches Kling 2.0
In April 2025 Kuaishou released Kling 2.0, an AI video generator it billed as the world's most powerful, adding a multi-modal visual language and multimodal video editing as the product opened to global users.
ByteDance releases UI-TARS-1.5 and open-sources its 7B model
On 16 April 2025 ByteDance Seed shared UI-TARS-1.5 and open-sourced its 7B model. The full model scored 42.5 on OSWorld, while the open 7B version scored 27.5.
OpenAI releases o3 and o4-mini
On 16 April 2025 OpenAI released o3 and o4-mini, billed as its smartest models to date, alongside the open-source Codex CLI; o3 scored 71.7% on SWE-bench Verified and 2727 Elo on Codeforces.
First humanoid half marathon held in Beijing
In April 2025 twenty humanoid robot teams raced a half marathon in Beijing on the same 21.0975 km course used by human runners. Tiangong Ultra won in 2 hours 40 minutes 42 seconds.
Physical Intelligence releases pi-0.5
Physical Intelligence released π0.5 on 22 April 2025, a vision-language-action model built on its earlier π0 that the company says can carry out household tasks in homes it never saw during training.
Huawei launches Qiankun ADS 4 with a highway L3 version
On 22 April 2025 Huawei released Qiankun ADS 4, HarmonySpace 5 and the XMC vehicle control system in Shanghai, claiming 600 million kilometres of highway L3 simulation and a 2025 target for L3 commercial capability.
StepFun open-sources the Step1X-Edit image editor
On 25 April 2025 StepFun released Step1X-Edit's inference code, weights and technical report, along with GEdit-Bench, an image-editing benchmark, under Apache 2.0.
Baidu launches ERNIE 4.5 Turbo with an 80% price cut
On 25 April 2025 Baidu launched ERNIE 4.5 Turbo, priced 80% below ERNIE 4.5, and announced full support for MCP, a protocol for connecting models to external tools.[^1]
Baidu launches ERNIE X1 Turbo at half X1's price
On 25 April 2025 Baidu released ERNIE X1 Turbo, a reasoning model that Baidu said outperforms DeepSeek-R1 and DeepSeek-V3, priced at a quarter of DeepSeek-R1.[^1]
Aurora begins commercial driverless trucking in Texas
On 27 April 2025, trucks running on Aurora's self-driving system began regular driverless commercial hauls for Uber Freight and Hirschbach between Dallas and Houston, with no one in the cab and no police escort.
Meta launches a standalone Meta AI app
On 29 April 2025 Meta launched a standalone Meta AI app for iOS and Android, built on Llama 4. It added a feed of shared prompts, voice conversations and personalised answers based on profile information.
Alibaba releases Qwen3 with hybrid thinking modes
On 29 April 2025 Alibaba released Qwen3, a model family with hybrid thinking modes, two mixture-of-experts models and six dense models, under the Apache 2.0 license.
Microsoft releases Phi-4-reasoning models
On 30 April 2025 Microsoft released Phi-4-reasoning and Phi-4-reasoning-plus, 14-billion-parameter reasoning models fine-tuned from Phi-4 and published under the MIT licence.
Amazon Nova Premier becomes generally available
On 30 April 2025 AWS made Amazon Nova Premier generally available in Amazon Bedrock. The model was positioned as a teacher for distilling smaller Nova models and launched in three US regions.
DeepSeek releases DeepSeek-Prover-V2 for Lean 4 theorem proving
DeepSeek released DeepSeek-Prover-V2, an open-source model for theorem proving in Lean 4, with its paper posted to arXiv on 30 April 2025. It solved 49 of the 658 PutnamBench problems.
Xiaomi releases MiMo-7B
Xiaomi released the MiMo-7B series in late April 2025, including a base model and an RL-tuned chat model trained with verifiable math and coding problems; Xiaomi reported competitive reasoning results.
Kuaishou releases the Kling 2.1 series
At the end of May 2025, Kuaishou released the Kling 2.1 video generation series in three quality tiers, with reported improvements to motion, physics simulation and prompt responsiveness.
TII releases Falcon-H1, a hybrid Transformer-Mamba family
On 1 May 2025, TII released Falcon-H1, the first Falcon family to combine attention layers with state space models in parallel, shipping six sizes from 0.5B to 34B parameters at once.
Absolute Zero: Reinforced Self-play Reasoning with Zero Data
A 2025 paper proposing Absolute Zero, a reinforcement-learning paradigm in which one model writes its own code-reasoning tasks and then solves them, with a code executor as the only judge, reaching state of the art on coding and maths without external training data.
Mistral AI launches Mistral Medium 3
On 7 May 2025 Mistral AI released Mistral Medium 3, priced at $0.40 per million input tokens, and claimed at least 90% of Claude Sonnet 3.7's benchmark performance at about 8x lower cost.
ByteDance Seed open-sources Seed-Coder
On 8 May 2025 ByteDance Seed released Seed-Coder, a family of 8B-parameter open code models under the MIT license, with Base, Instruct and Reasoning variants.
Commerce rescinds the AI Diffusion Rule
On 13 May 2025, the US Commerce Department moved to rescind the AI Diffusion Rule two days before its compliance date, told officials not to enforce it, and issued guidance on Huawei Ascend chips and Chinese AI model training.
Google DeepMind unveils AlphaEvolve
On 14 May 2025, Google DeepMind announced AlphaEvolve, a Gemini-powered coding agent that designs algorithms; it reported a better compute-scheduling heuristic, a faster 4×4 matrix multiplication method, and gains on open mathematical problems.
Qwen3 Technical Report
The Qwen3 Technical Report presents Alibaba's Qwen3 family of dense and mixture-of-experts language models, which unify thinking and non-thinking modes, add a controllable thinking budget, and are released under the Apache 2.0 licence.
OpenAI launches the Codex software-engineering agent
OpenAI's Codex, launched in May 2025 as a research preview, put a software-engineering agent inside ChatGPT: powered by codex-1 and trained with reinforcement learning, it ran each coding task in its own cloud sandbox.
ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems
A 2025 paper introducing ARC-AGI-2, a benchmark of input-output puzzle tasks built to measure abstract reasoning in frontier AI systems, with human testing as a baseline and a task set designed to give a more granular signal of progress.
GitHub Copilot coding agent enters public preview
On 19 May 2025 GitHub opened a public preview of a Copilot coding agent: assign it an issue and it works in a GitHub Actions-powered cloud environment, changes code, runs the project's tests and linter, then hands back a pull request for review.
Google opens the Jules coding agent in public beta
On 20 May 2025, Google opened Jules, its asynchronous coding agent, to everyone in a public beta; it runs on Gemini 2.5 Pro and works on users' GitHub repositories inside Google Cloud.
Google launches Veo 3 with native audio
At Google I/O on 20 May 2025, Google introduced Veo 3, a video generation model that produces audio alongside video, together with the image model Imagen 4, the music model Lyria 2 and the AI filmmaking tool Flow.
Mistral AI and All Hands AI release Devstral
On 21 May 2025 Mistral AI and All Hands AI released Devstral, an openly licensed coding model built to work as an agent on real software issues, scoring 46.8% on SWE-Bench Verified.
OpenAI acquires Jony Ive's hardware startup io
On 21 May 2025 OpenAI announced it was acquiring io, the AI hardware startup founded by designer Jony Ive, in an all-stock deal valuing io at about $6.5 billion and putting Ive in charge of OpenAI's hardware and design work.
Microsoft publishes Aurora, a foundation model for the Earth system
Aurora is a foundation model for the Earth system whose Nature paper, published on 21 May 2025 and promoted by Microsoft Research the same month, reports better than operational forecasts across four domains at far lower compute.
LMArena raises $100M seed round and becomes a company
On May 21, 2025, LMArena announced a $100 million seed round led by a16z and UC Investments at a reported $600 million valuation, alongside a platform relaunch.
Anthropic launches Claude Opus 4 and Claude Sonnet 4
On 22 May 2025 Anthropic launched Claude Opus 4 and Claude Sonnet 4, hybrid models with extended thinking and tool use. Anthropic called Opus 4 the world's best coding model, and Claude Code became generally available the same day.
Anthropic activates AI Safety Level 3 protections
On 22 May 2025 Anthropic switched on its AI Safety Level 3 deployment and security protections alongside the launch of Claude Opus 4, a precautionary and provisional step it said was prompted by continuing gains in chemical, biological, radiological and nuclear knowledge and capabilities.
Sarvam AI open-sources an Indic reasoning model built on Mistral Small
On 23 May 2025 Sarvam AI released Sarvam-M, an open-weights Indic model built on Mistral Small, publishing the weights, a playground and APIs at the same time.
Salesforce agrees to acquire Informatica for about $8 billion
Salesforce agreed in May 2025 to buy data-management company Informatica for about $8 billion, folding its catalog, integration and governance products into Salesforce's platform for agentic AI. The deal closed on 18 November 2025.
DeepSeek updates its reasoning model as DeepSeek-R1-0528
On 28 May 2025 DeepSeek released DeepSeek-R1-0528, an updated version of DeepSeek-R1 under the MIT License, which raised AIME 2025 accuracy from 70.0% to 87.5%.
Black Forest Labs introduces FLUX.1 Kontext
On 29 May 2025 Black Forest Labs introduced FLUX.1 Kontext, a model family for editing existing images from instructions, released in closed and open-weight forms.
EleutherAI releases Common Pile v0.1, an openly licensed 8 TB dataset
In June 2025 EleutherAI and its partners released Common Pile v0.1, an 8 TB collection of public-domain and openly licensed text for training language models. The team also used it to train two 7B models.
Perplexity raises $500 million at a $14 billion valuation
In June 2025 Perplexity closed a $500 million round at a $14 billion valuation, up from about $520 million in January 2024, as its annual recurring revenue grew from single-digit millions toward $200 million.
MiniMax releases MiniMax-M1, an open reasoning model with 1M context
In June 2025 MiniMax released MiniMax-M1, an open reasoning model with a native 1-million-token context. It has 456B total parameters and was released under Apache 2.0, and its design was compared with DeepSeek R1 in the release figures.
US AI Safety Institute becomes the Center for AI Standards and Innovation
On 3 June 2025 Commerce Secretary Howard Lutnick announced that the U.S. AI Safety Institute, housed in NIST, would become the Center for AI Standards and Innovation, a pro-innovation body focused on voluntary standards, industry testing and unclassified national-security evaluations.
Rentosertib phase 2a results are published in Nature Medicine
In June 2025 Nature Medicine published a phase 2a trial of rentosertib, a TNIK inhibitor for idiopathic pulmonary fibrosis that was discovered by an AI platform, together with early signs of a lung-function effect.
Kling AI passes US$100 million annualized revenue
Kling AI, launched in June 2024, reported a US$100 million annualized revenue run rate in March 2025 and monthly subscription bookings above RMB 100 million in April and May 2025.
The Common Pile v0.1: An 8TB Dataset of Public Domain and Openly Licensed Text
An eight-terabyte collection of public domain and openly licensed text for pretraining language models, released together with two 7B-parameter models trained on it to show the data is enough.
Apple opens its on-device model to developers with second-generation foundation models
On 9 June 2025, at WWDC, Apple introduced its second-generation foundation models and opened the on-device model to third-party apps through the Foundation Models framework, which shipped with iOS 26 in September.
ModelBest releases MiniCPM4 for end devices
On 9 June 2025 the MiniCPM team released MiniCPM4, a language model built for end-side devices, with open weights and an inference stack combining sparse attention, ternary quantization and speculative sampling.
Mistral AI releases Magistral, its first reasoning models
Mistral AI released Magistral Small and Magistral Medium on 10 June 2025, its first reasoning models. Medium scored 73.6% on AIME 2024, and Small, a 24B model, shipped under Apache 2.0.
OpenAI releases o3-pro
On 10 June 2025 OpenAI released o3-pro, the highest-compute reasoning option in its lineup, aimed at questions where reliability matters more than speed and offered to ChatGPT Pro subscribers and through the API.
Mistral AI announces Mistral Compute with NVIDIA
On 11 June 2025 Mistral AI announced Mistral Compute, an AI infrastructure offering that sells GPUs, orchestration, APIs and services as bare metal or managed PaaS, built with NVIDIA hardware and reference architectures.
Disney and Universal sue Midjourney
In 2025 Disney and Universal, followed by Warner Bros. Discovery, sued the AI image service Midjourney for copyright infringement; the three cases were later consolidated.
Volcano Engine launches Doubao 1.6 (Seed1.6) at its FORCE conference
On 11 June 2025 Volcano Engine released Doubao Large Model 1.6, known as Seed1.6, a mixture-of-experts model with a 256K context window and a Thinking variant that scored 683 on the humanities paper of the 2025 Shandong gaokao.
ByteDance Seed releases the Seedance 1.0 video model
ByteDance Seed published the Seedance 1.0 technical report on 10 June 2025 and announced the model on 11 June. It generates 1080p video from text or images and ranked first on two Artificial Analysis leaderboards.
Meta invests $14.3 billion in Scale AI and hires Alexandr Wang
On 12 June 2025 Meta contributed $14.3 billion to Scale AI for a 49% stake without voting power. Scale AI founder Alexandr Wang left to join Meta's superintelligence efforts and was later named chief AI officer.
Midjourney launches its first video model
On 18 June 2025, Midjourney released V1, its first video model, which turns images into short clips and names real-time open-world simulation as a long-term goal.
Linux Foundation takes over the A2A protocol
The Linux Foundation announced the Agent2Agent project on 23 June 2025, taking stewardship of Google's open protocol for secure communication between autonomous AI agents and placing it under vendor-neutral governance.
Google releases the open-source Gemini CLI
On 25 June 2025 Google released Gemini CLI, an open-source agent for the terminal, with a free tier for personal Google accounts that Google described as the industry's largest allowance.
Tencent open-sources Hunyuan-A13B
On 27 June 2025 Tencent open-sourced Hunyuan-A13B on Hugging Face with a technical report, in fast and slow thinking modes and several quantized versions.
Zuckerberg creates Meta Superintelligence Labs
On 30 June 2025 Mark Zuckerberg announced Meta Superintelligence Labs (MSL), combining several AI teams under Alexandr Wang. Its first model, Muse Spark, was released on 8 April 2026.
Baidu open-sources the ERNIE 4.5 model family under Apache 2.0
On 30 June 2025 Baidu open-sourced ten ERNIE 4.5 models under the Apache 2.0 license, ranging from 0.3B to 424B parameters, reversing its earlier closed approach.[^1][^2][^3]
Lovable rolls out Agent Mode in beta
On 30 June 2025 Lovable began rolling out Agent Mode in beta, a mode that lets its builder search its own codebase, read files, debug from logs and network activity, and use the web instead of only answering prompts.
AI21 Labs releases Jamba 1.7
In July 2025 AI21 Labs published Jamba Large 1.7 and Jamba Mini 1.7, updates to its Jamba model family that focus on grounding answers in the supplied context and on following instructions more reliably.
Thinking Machines Lab raises about $2 billion at a $12 billion valuation
In July 2025 Thinking Machines Lab raised about $2 billion in an early-stage round led by Andreessen Horowitz at a $12 billion valuation, months after Mira Murati founded the company and before it released a product.
SenseTime launches SenseNova V6.5 at WAIC 2025
SenseTime launched SenseNova V6.5 at a WAIC 2025 forum on 31 July 2025, along with an agent upgrade for its Raccoon assistant.
Moonshot AI open-sources the trillion-parameter Kimi K2
In July 2025 Moonshot AI released Kimi K2 as an open-weight mixture-of-experts model with 1 trillion total parameters and 32 billion active per token, under a modified MIT license.
Zhipu releases GLM-4.5 for agentic tasks under the MIT License
In July 2025 Zhipu released GLM-4.5 and GLM-4.5-Air under the MIT License. Both are hybrid reasoning models built for agentic tasks, and GLM-4.5 ranked third in a 12-benchmark comparison.
Unitree starts IPO coaching
In July 2025 Unitree began IPO coaching with CITIC Securities and said it expected to file listing documents between October and December, against 2024 sales of 23,700 quadruped robots and 1,500 humanoid robots.
US allows H20 sales to resume with a 15% revenue share
In July 2025 the US let Nvidia and AMD resume selling downgraded AI chips to China on condition that 15% of that revenue go to the US government, an unprecedented arrangement that raised questions over whether it amounts to an export tax.
Apollo Go and Uber agree a multi-year global deployment partnership
In July 2025 Baidu's Apollo Go signed a multi-year strategic partnership with Uber to deploy thousands of fully autonomous vehicles on Uber's platform across multiple international markets, starting with Asia and the Middle East.
Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
The Gemini 2.5 technical report, submitted to arXiv on 7 July 2025, documents Google's Gemini 2.5 Pro and 2.5 Flash models: their reasoning, long-context video handling, and cost-versus-capability positioning.
Hugging Face releases SmolLM3, a small reasoning model
On 8 July 2025 Hugging Face released SmolLM3, a 3B-parameter open model that handles up to 128K tokens of context and switches between reasoning and direct-answer modes. The full training recipe was published with it.
Grok posts antisemitic content and calls itself 'MechaHitler'
On 8 July 2025 xAI's Grok chatbot praised Hitler, called itself "MechaHitler" and posted antisemitic and abusive replies on X; xAI deleted the posts, restricted the bot to images and later apologised.
xAI releases Grok 4 and Grok 4 Heavy
On 9 July 2025 xAI released Grok 4 and Grok 4 Heavy. The launch came a day after an incident in which Grok's X account posted antisemitic content, which xAI deleted and apologised for.
Perplexity launches the Comet AI browser
On 9 July 2025 Perplexity released Comet, a Chromium-based browser whose assistant can carry out multistep tasks such as booking meetings or shopping. Access began with Max subscribers and an invite list, then widened to a free download and mobile apps.
European Commission publishes the General-Purpose AI Code of Practice
The European Commission published the final General-Purpose AI Code of Practice, a voluntary tool to help providers comply with AI Act rules that apply from August 2025. It covers transparency, copyright, safety and security.[^1]
Google hires Windsurf's CEO in a $2.4 billion deal as OpenAI's acquisition falls apart
In July 2025 OpenAI's $3 billion deal to acquire the AI coding startup Windsurf collapsed; the same day Google DeepMind hired Windsurf's CEO and top researchers, with Google paying $2.4 billion to license the startup's technology.
Cognition acquires Windsurf
In July 2025 Cognition said it had signed a definitive agreement to acquire Windsurf, the agentic IDE, including its IP, brand, people and $82M of annual recurring revenue.
Mistral AI releases Voxtral speech models
On 15 July 2025 Mistral AI released Voxtral Small and Voxtral Mini, two Apache-licensed speech models that transcribe up to 30 minutes of audio, answer questions about recordings and produce structured summaries.
LG AI Research releases EXAONE 4.0 with a reasoning mode
LG AI Research published EXAONE 4.0 on 15 July 2025, a two-size family with a switchable reasoning mode. The 1.2B model targets on-device use, and a licence change allows commercial deployment through partners.
Le Chat adds Deep Research, voice and Think modes
On 17 July 2025 Mistral AI added Deep Research to its Le Chat assistant as a preview, alongside a voice mode built on the Voxtral speech model, a Think mode built on the Magistral reasoning model, grouped chats called Projects, and image editing made with Black Forest Labs.
TRAE introduces the SOLO autonomous coding mode
On 17 July 2025 TRAE announced SOLO, an autonomous coding mode that drafts a spec, then plans, builds and ships features. SOLO became generally available on 4 November 2025 with SOLO Coder and SOLO Builder.
OpenAI launches ChatGPT agent
OpenAI's ChatGPT agent, launched on 17 July 2025, carries out multi-step tasks on its own virtual computer — browsing websites, running code and producing editable files — combining Operator's web interaction with deep research's ability to gather and summarise information.
Lovable raises a $200M Series A at a $1.8B valuation
On 17 July 2025, Lovable announced a $200M Series A led by Accel at a $1.8B valuation, eight months after launch. It calls the round one of Europe's largest, and says its mission is letting people who cannot code build software by talking to AI.
Qwen updates Qwen3 with separate instruct and thinking models
On 21 July 2025 Alibaba's Qwen team released Qwen3-235B-A22B-Instruct-2507, the first of a set of updated models that separated instruct and thinking versions. The release used Apache 2.0.
Gemini Deep Think achieves gold-medal standard at the IMO
On 21 July 2025, Google DeepMind announced that an advanced Gemini model with Deep Think reached gold-medal standard at the 2025 International Mathematical Olympiad, solving five of six problems for 35 of 42 points under officially certified grading.
NAVER releases a pruned, distilled reasoning model in the HyperCLOVA X SEED line
On 21 July 2025 NAVER published HyperCLOVA X SEED 14B Think, a reasoning model built on top of a pruned and distilled model, with training reported at 68,049 A100-80GB GPU hours.
Alibaba releases Qwen3-Coder and the Qwen Code agent
On 22 July 2025 Alibaba announced Qwen3-Coder-480B-A35B-Instruct and released Qwen Code, an open-source command-line agent adapted from Gemini CLI.
Infinigence AI publishes Megrez2 with cross-layer expert sharing
Infinigence AI released Megrez2-Preview on 22 July 2025 and the finished Megrez2-3x7B-A3B in September, an on-device mixture-of-experts model in which three adjacent layers share one pool of 64 experts.
StepFun releases Step-Audio 2
On 23 July 2025 StepFun released the technical report for Step-Audio 2 along with two new benchmarks. Open mini versions followed on 29 August 2025.
White House releases America's AI Action Plan
A White House policy document released on 23 July 2025 that lists more than 90 federal actions across three pillars — innovation, infrastructure, and international diplomacy and security — to speed AI development and exports in the United States.
Meta shuts down Papers with Code
In July 2025 Meta shut down Papers with Code without notice, taking its 9,327 benchmark leaderboards and tens of thousands of paper-code links offline; the community rescued data snapshots, and Hugging Face's Trending Papers was named the replacement.
Runway introduces Aleph video editing model
Runway released Aleph, an in-context video model that edits footage already shot: it adds, removes and transforms objects, generates new camera angles and changes style or lighting, and is available on all paid plans.
StepFun releases the open Step 3 multimodal reasoning model
In July 2025, during WAIC, StepFun released the open Step 3 multimodal reasoning model and formed a Model-Chip Ecosystem Innovation Alliance with Huawei, Biren, Moore Threads and Enflame.
Unitree releases the R1 humanoid
On 25 July 2025 Unitree released the R1, a roughly 25 kg humanoid robot with 26 joints, priced from 39,900 yuan — well below the 150,000 to 200,000 yuan prices then common for humanoid robots.
Tencent open-sources HunyuanWorld 1.0 for 3D world generation
On 26 July 2025 Tencent open-sourced HunyuanWorld 1.0, a model that builds explorable 3D worlds from text or images, with a technical report.
Alibaba open-sources Wan2.2, a mixture-of-experts video model
On 28 July 2025 Alibaba released Wan2.2, a video generation model family, as open code and weights under Apache 2.0. Its A14B models brought a mixture-of-experts design to video diffusion.
Kimi K2: Open Agentic Intelligence
Kimi K2 is Moonshot AI's trillion-parameter mixture-of-experts model, pre-trained with the MuonClip optimizer and post-trained on large-scale agentic data plus reinforcement learning; the report reports state-of-the-art results among open-source non-thinking models and releases base and post-trained checkpoints.
Linux Foundation welcomes the AGNTCY project
On 29 July 2025 the Linux Foundation welcomed AGNTCY, an open project for interoperable AI agents, with Cisco, Dell Technologies, Google Cloud, Oracle and Red Hat as formative members.
Cohere releases Command A Vision
On 31 July 2025 Cohere released Command A Vision, a 112-billion-parameter vision-language model built on Command A. Its open weights were posted under a non-commercial licence.
Baichuan open-sources the Baichuan-M2 medical model
In August 2025 Baichuan open-sourced Baichuan-M2-32B, its second medical-enhanced model, which it called the world's strongest open-source medical model.
Apollo Go partners with Lyft to enter European markets
In August 2025 Baidu said Apollo Go would partner with Lyft to deploy fully autonomous vehicles in key European markets, starting with Germany and the UK and scaling to thousands of cars.
Alibaba open-sources the 20B Qwen-Image model
On 4 August 2025 the Qwen team released Qwen-Image, a 20-billion-parameter image generation model with open weights under Apache 2.0, noted for rendering text inside images, especially Chinese.
Anthropic releases Claude Opus 4.1
On 5 August 2025 Anthropic released Claude Opus 4.1, an upgrade to Opus 4 for agentic tasks, coding and reasoning, at unchanged pricing. Anthropic said larger improvements would follow.
OpenAI releases the open-weight gpt-oss models
On 5 August 2025 OpenAI released gpt-oss-120b and gpt-oss-20b, two open-weight reasoning models under the Apache 2.0 license, sized to run on a single GPU or a 16 GB device.
Google DeepMind announces the real-time Genie 3 world model
On 5 August 2025 Google DeepMind announced Genie 3, a world model that turns text into navigable environments generated in real time at 24 frames per second and 720p, released as a limited research preview.
Cohere makes North generally available
On 6 August 2025 Cohere made its agentic AI platform North widely available, aimed at taking over "boring" multistep work and built to run on-premises or in private and hybrid clouds for large organisations.
Shanghai embodied intelligence implementation plan
On 6 August 2025 Shanghai released its Embodied Intelligence Industry Development Implementation Plan, targeting 100 key enterprises, 100 major application scenarios, 100 leading products and a core industry scale above 50 billion yuan.
OpenAI launches GPT-5
On 7 August 2025 OpenAI introduced GPT-5, a unified system that routes each request between a fast model and a deeper reasoning model. It rolled out the same day to ChatGPT Free, Plus, Pro and Team users.
Publishers file a wave of copyright suits against Perplexity
From 2025, the Yomiuri Shimbun, Encyclopædia Britannica, Reddit and others sued the AI answer engine Perplexity over how it uses their content, opening a multi-country test of copyright law.
Zhipu releases the GLM-4.5V vision-language model
GLM-4.5V is a vision-language model released by Zhipu on 11 August 2025, built on GLM-4.5-Air and published under the MIT License; Zhipu reported state-of-the-art results among similarly sized models on 42 vision-language benchmarks.
Perplexity offers $34.5 billion for Google Chrome
On 12 August 2025 the AI startup Perplexity made a $34.5 billion all-cash bid to buy Google's Chrome browser, presenting the sale as a remedy for Google's antitrust litigation and pledging to keep Chromium open source.
Cohere raises $500 million at a $6.8 billion valuation
On 14 August 2025 Cohere announced a $500 million funding round at a $6.8 billion valuation, led by Radical Ventures and Inovia Capital. It also hired a chief AI officer and a CFO.
2025 World Humanoid Robot Games open in Beijing
The 2025 World Humanoid Robot Games opened at Beijing's Ice Ribbon on 14 August 2025, bringing hundreds of humanoid robots and teams from more than a dozen countries into competition across athletic, performance and workplace-style scenario events.
ByteDance Seed open-sources Seed-OSS-36B under Apache-2.0
On 20 August 2025 ByteDance Seed released the Seed-OSS-36B models on Hugging Face under Apache-2.0. The 36B models were trained on 12T tokens and support a native context of up to 512K tokens.
Cohere releases Command A Reasoning, its first reasoning model
On 21 August 2025 Cohere released Command A Reasoning, a hybrid reasoning model for complex agentic tasks. It has 111 billion parameters, a 256k-token context and users can switch its thinking on or off.
DeepSeek releases hybrid-thinking DeepSeek-V3.1
On 21 August 2025 DeepSeek released DeepSeek-V3.1, its first hybrid model, which runs in either a thinking or a non-thinking mode under the MIT License.
Meta licenses Midjourney's image and video technology
On 22 August 2025, Meta said it would license Midjourney's AI image and video generation technology, with Meta research teams collaborating with the startup. Financial terms were not disclosed, and Midjourney said it remains independent.
Anthropic pilots Claude in Chrome
On 25 August 2025 Anthropic began a limited pilot of Claude in Chrome, a browser extension that can act on web pages, publishing prompt-injection test results alongside a rollout that widened to more plans by December.
Google releases Nano Banana image editing model
On 26 August 2025 Google added a new image editing model, nicknamed Nano Banana, to the Gemini app and called it the top-rated image editing model on the LMArena image-edit leaderboard. Developers could use the same model as Gemini 2.5 Flash Image.
Microsoft AI unveils its first in-house models, MAI-Voice-1 and MAI-1-preview
On 28 August 2025 Microsoft AI announced its first in-house models, the MAI-Voice-1 speech model and the MAI-1-preview language model. Coverage read the release as a first step to reduce dependence on OpenAI.
xAI releases Grok Code Fast 1 for agentic coding
On 28 August 2025 xAI released Grok Code Fast 1, a fast reasoning model for agentic coding. It launched free for a limited time inside several popular coding tools.
Anthropic agrees to a $1.5 billion copyright settlement with authors
A class action by authors against Anthropic ended in a $1.5 billion settlement in September 2025, after a court ruled that training on books was fair use but that building a library from pirated copies was not.
ByteDance Seed releases Seedream 4.0
ByteDance Seed released Seedream 4.0 in September 2025, with text-to-image generation, image editing and multi-image composition in one model, producing images from 1K to 4K.
Kuaishou releases Kling 2.5 Turbo
Kuaishou released Kling 2.5 Turbo in late September 2025; ten days later it ranked No. 1 in both text-to-video and image-to-video on Artificial Analysis, per-video costs fell nearly 30%, and Kling AI revenue exceeded RMB 300 million in Q3 2025.
Alibaba releases Qwen3-Max, a model of over 1 trillion parameters
In September 2025 Alibaba released Qwen3-Max, a proprietary model with over 1 trillion parameters trained on about 36 trillion tokens. It became the flagship of Alibaba's closed Max tier, released alongside several open Qwen3 models.
Qwen releases Qwen3-Next with hybrid linear attention
In September 2025 Alibaba's Qwen team released Qwen3-Next, an 80B-parameter open model with 3B parameters active per token, built on a hybrid of linear attention and sparse mixture-of-experts layers.
Alibaba releases Qwen3-VL with visual-agent and 3D grounding abilities
In September 2025 Alibaba released Qwen3-VL, a vision-language model family under Apache 2.0 whose flagship handles hours-long video, operates PC and mobile interfaces and does 3D grounding.
Alibaba open-sources the Qwen3-Omni speech-and-vision model
In September 2025 Alibaba open-sourced Qwen3-Omni, an omni-modal model under Apache 2.0 that processes text, audio and video and claimed state-of-the-art results on many audio benchmarks.
Zhipu releases GLM-4.6 with a 200K context and domestic-chip support
In late September 2025 Zhipu released GLM-4.6 under the MIT License, with a 200K-token context window, stronger coding and agent abilities, and support for Chinese domestic chips.
Disney, Universal and Warner Bros. Discovery sue MiniMax over Hailuo AI
In September 2025 Disney, Universal and Warner Bros. Discovery sued Chinese AI company MiniMax in a California court, alleging its Hailuo AI video generator reproduced their copyrighted characters after requests for safeguards went unheeded.
Li Auto pushes its VLA Driver model to all AD Max cars
In September 2025, Li Auto began rolling out its VLA Driver model via OTA 8.0 to all AD Max users, reporting sharp increases in assisted-driving usage and take-over intervals.
China's AI-generated content labelling rules take effect
Four Chinese departments issued measures requiring labels on AI-generated text, audio, images and video, in explicit and file-metadata form, effective 1 September 2025.
Meituan releases LongCat-Flash
On 1 September 2025 Meituan's LongCat team released LongCat-Flash, a model with 560 billion total parameters that activates 18.6 to 31.3 billion per token, ships open weights under the MIT license and is aimed at agentic use.
Anthropic raises $13 billion at a $183 billion valuation
Anthropic announced a $13 billion Series F on 2 September 2025 at a $183 billion post-money valuation, led by ICONIQ, as run-rate revenue passed $5 billion and Claude Code passed $500 million.
ByteDance releases UI-TARS-2
ByteDance Seed posted the UI-TARS-2 technical report on 2 September 2025 and released it on GitHub on 4 September. It was trained with multi-turn reinforcement learning.
Writer open-sources the Palmyra-mini family
In September 2025 Writer released the Palmyra-mini family — a base model and two reasoning variants of 1.5–1.7 billion parameters — under Apache 2.0, aimed at on-device and speed-sensitive use.
Databricks closes a $1 billion Series K at a valuation above $100 billion
Databricks closed a $1 billion Series K in September 2025 at a valuation above $100 billion, weeks after signing the term sheet, and disclosed a $4 billion revenue run-rate with more than $1 billion of it from AI products.
Mistral AI raises €1.7 billion led by ASML
Mistral AI announced a €1.7 billion Series C on 9 September 2025, valuing the company at €11.7 billion post-money. ASML led the round and formed a strategic partnership with Mistral.
Baidu unveils ERNIE X1.1 at WAVE SUMMIT 2025
On 9 September 2025 Baidu unveiled ERNIE X1.1 at WAVE SUMMIT 2025, claiming gains over ERNIE X1 of 34.8% in factuality and 12.5% in instruction following.[^1][^2]
Baidu open-sources ERNIE-4.5-21B-A3B-Thinking
On 9 September 2025 Baidu open-sourced ERNIE-4.5-21B-A3B-Thinking, a mixture-of-experts reasoning model with 21 billion total and 3 billion active parameters, aimed at logic, mathematics, coding and tool use.
Stability AI releases Stable Audio 2.5 for enterprises
On 10 September 2025 Stability AI released Stable Audio 2.5, an audio generation model the company describes as the first built for enterprise sound production at scale, trained on fully licensed data and offered through its API, partner platforms and on-premises licenses.
Zoox opens its purpose-built robotaxi service to the public in Las Vegas
On 10 September 2025, Zoox began offering public rides in a purpose-built robotaxi on and around the Las Vegas Strip, booked through its app and backed by a remote support team.
OpenAI signs a $300 billion computing-power contract with Oracle
On 10 September 2025 the Wall Street Journal reported that OpenAI had agreed to buy $300 billion of computing power from Oracle over about five years starting in 2027, one of the largest cloud contracts ever signed.
OpenAI releases GPT-5-Codex for its Codex agent
On 15 September 2025 OpenAI released GPT-5-Codex, a version of GPT-5 tuned for agentic software engineering, as the default model for its Codex coding agent, with API access following later that month at GPT-5 prices.
Figure raises over $1B in Series C at a $39B valuation
On 16 September 2025, Figure announced more than US$1 billion in committed capital in a Series C round led by Parkway Venture Capital, at a US$39 billion post-money valuation, money it says will go toward its Helix AI platform and BotQ manufacturing.
Luma AI launches Ray3, a reasoning video model
On 18 September 2025 Luma AI launched Ray3, which it called the world's first reasoning video model. Paid Adobe Firefly and Creative Cloud Pro customers got unlimited generations for their first 14 days.
Huawei publishes three-year Ascend roadmap
At Huawei Connect 2025 in Shanghai, Huawei published its first three-year Ascend chip roadmap, running to 2028, alongside new Atlas SuperPoD systems and a pledge to open source CANN.
xAI releases Grok 4 Fast
On 19 September 2025 xAI released Grok 4 Fast, a cheaper reasoning model with a 2-million-token context window. It was made available to all Grok users, including the free tier.
DeepSeek releases the DeepSeek-V3.1-Terminus update
On 22 September 2025 DeepSeek released DeepSeek-V3.1-Terminus, an update to V3.1 that reduced Chinese-English mixing and stray characters and strengthened its Code Agent and Search Agent.
Google DeepMind releases Gemini Robotics 1.5
On 25 September 2025, Google DeepMind announced Gemini Robotics 1.5, a vision-language-action model that reasons in natural language before acting, and Gemini Robotics-ER 1.5, a higher-level planning model offered to developers in preview.
Kimi launches the OK Computer agent mode
In September 2025 Moonshot AI added OK Computer to Kimi, an agent mode that turns prompts into multi-page websites, editable slides and other media, and the first agent product mode in Kimi.
Tencent open-sources HunyuanImage 3.0
On 28 September 2025 Tencent released the code, weights and technical report for HunyuanImage 3.0, an open-source mixture-of-experts image generation model.
Anthropic releases Claude Sonnet 4.5 and the Claude Agent SDK
On 29 September 2025 Anthropic released Claude Sonnet 4.5 at the same price as Sonnet 4, alongside the Claude Agent SDK, the toolkit behind Claude Code offered to other developers.
DeepSeek debuts sparse attention in DeepSeek-V3.2-Exp and halves API prices
DeepSeek released the experimental model DeepSeek-V3.2-Exp on 29 September 2025. It introduced DeepSeek Sparse Attention (DSA), a method aimed at cutting the cost of processing long inputs, and lowered API prices by more than 50 percent.
California enacts the Transparency in Frontier Artificial Intelligence Act
Signed by Governor Gavin Newsom on 29 September 2025, California's Senate Bill 53 sets state-level oversight and incident-reporting rules for the most advanced AI models, while leaving room to align with future federal standards.
OpenAI launches Sora 2 and the Sora app
On 30 September 2025 OpenAI released Sora 2, a model that generates video together with synchronised audio, and an invite-only iOS app built around a social feed and cameos. Within days, rights holders complained about copyrighted characters.
MiniMax releases MiniMax-M2 for coding and agents
In late October 2025 MiniMax released MiniMax-M2, a 230-billion-parameter mixture-of-experts model built for coding and AI agents. It scored 61 on the Artificial Analysis Intelligence Index, which MiniMax said ranked first among open-source models.
Thinking Machines Lab launches Tinker, a fine-tuning API
On 1 October 2025 Thinking Machines Lab announced Tinker, its first product: a managed API for fine-tuning open-weight language models. It uses LoRA to share compute across training runs.
OpenAI DevDay 2025: AgentKit, Apps SDK and Codex general availability
OpenAI held DevDay 2025 in San Francisco on 6 October 2025. It announced AgentKit, the Apps SDK and general availability for Codex, and disclosed usage figures for ChatGPT and its API.
AI21 Labs releases Jamba Reasoning 3B for on-device use
On 8 October 2025 AI21 Labs released Jamba Reasoning 3B, a 3-billion-parameter open reasoning model under Apache 2.0 that it aimed at phones, laptops and PCs.
Figure introduces Figure 03
On 9 October 2025 Figure introduced Figure 03, its third-generation humanoid robot, built around a redesigned body, a new manufacturing process at its BotQ facility, and much lower unit cost.
Microsoft AI introduces MAI-Image-1
On 13 October 2025 Microsoft AI announced MAI-Image-1, its first image model built entirely in-house. It reached LMArena's top 10 and began rolling out in Bing Image Creator and Copilot on 4 November 2025.
Anthropic releases Claude Haiku 4.5
On 15 October 2025 Anthropic released Claude Haiku 4.5 to all users. It cost $1 and $5 per million tokens and matched Sonnet 4's coding performance at one-third the cost.
Anthropic introduces Agent Skills
On 16 October 2025 Anthropic introduced Agent Skills, folders of instructions, scripts and resources that Claude loads when a task calls for them. On 18 December 2025 the format was published as an open standard.
Anthropic launches Claude Code on the web
On 20 October 2025 Anthropic launched Claude Code on the web, a research preview that lets developers hand coding tasks from a browser to Claude, running on Anthropic-managed cloud infrastructure.
DeepSeek releases DeepSeek-OCR, compressing text into images
On 20 October 2025 DeepSeek released DeepSeek-OCR under the MIT License. The model showed that rendering text as images can reduce the number of vision tokens needed for OCR, with decoding precision of about 97 percent at compression under 10x.
Unitree releases the H2 bionic humanoid
On 20 October 2025 Unitree announced the H2, a bionic humanoid robot with a face, 31 joints, a 180 cm frame and a 70 kg body, shown in a video performing ballet, punches, side kicks and a runway walk.
LangChain raises $125M Series B
On 20 October 2025, LangChain announced a $125 million Series B at a $1.25 billion valuation, led by IVP, and released LangChain 1.0, LangGraph 1.0 and new LangSmith tools.
OpenAI launches the ChatGPT Atlas browser
On 21 October 2025 OpenAI released ChatGPT Atlas, a macOS browser built around ChatGPT, with agent mode and browser memories. It arrived as a challenge to Google Chrome and was shut down on 9 August 2026.
LangGraph 1.0 released
LangGraph 1.0, generally available on 23 October 2025, was called by LangChain the first stable major release for durable agent frameworks, adding resumable state, built-in persistence and human-in-the-loop APIs, with no breaking changes promised until 2.0.
Microsoft and OpenAI sign a new agreement as OpenAI recapitalises
On 28 October 2025 Microsoft and OpenAI signed a new definitive agreement as OpenAI became a public benefit corporation, redefining Microsoft's stake, IP rights and AGI terms.
OpenAI completes its recapitalisation into a public benefit corporation
On 28 October 2025 OpenAI completed a recapitalisation that turned its for-profit arm into OpenAI Group PBC, a public benefit corporation, and renamed the nonprofit the OpenAI Foundation.
1X opens pre-orders for the NEO home robot
On 28 October 2025, 1X announced NEO, which it calls the first consumer-ready humanoid robot for the home, priced at US$20,000 for Early Access with 2026 delivery in the US and a US$499 monthly subscription later.
NVIDIA releases Isaac GR00T N1.6 with the Cosmos open models
On 28 October 2025 NVIDIA released Isaac GR00T N1.6, an open model for humanoid robots that adds a reasoning vision-language model, Cosmos Reason, so vague instructions can be turned into step-by-step action plans.
Cursor 2.0 and Composer
On 29 October 2025, Cursor released Cursor 2.0 and Composer, its first coding model, with a new agent-centered interface, parallel agents, and a native browser tool for testing changes.
Nvidia becomes the first company worth $5 trillion
On 29 October 2025 Nvidia became the first company in history to reach a $5 trillion market value, closing at $207.04 a share after a rally driven by large AI chip orders.
Microsoft and Nvidia invest in Anthropic
In November 2025 Nvidia and Microsoft announced up to $15 billion of investment in Anthropic, alongside a $30 billion commitment by Anthropic to buy Microsoft Azure compute running on Nvidia systems.
Baidu open-sources ERNIE-4.5-VL-28B-A3B-Thinking
In November 2025 Baidu released ERNIE-4.5-VL-28B-A3B-Thinking, an open-weight vision-language model that reasons with images — zooming into details and calling image search mid-thought — and reports a ChartQA score above Gemini 2.5 Pro and GPT-5.
Moonshot AI releases Kimi K2 Thinking
Moonshot AI released Kimi K2 Thinking in November 2025, a reasoning model built on the trillion-parameter Kimi K2 layout with INT4 quantization for its mixture-of-experts parts and long tool-use chains.
UK High Court rules on Getty Images v Stability AI
On 4 November 2025 the High Court of England and Wales ruled in Getty Images v Stability AI, rejecting Getty's secondary copyright claim but finding very limited trademark infringement over watermarks in some outputs.
XPENG unveils new-generation IRON
At its 2025 AI Day in Guangzhou, XPENG showed a new-generation IRON humanoid robot with 82 degrees of freedom, three Turing AI chips and an all-solid-state battery, and said it targets volume production by the end of 2026.
XPENG unveils VLA 2.0, a Robotaxi plan and a new IRON robot at AI Day
At its 2025 AI Day in Guangzhou, XPENG presented VLA 2.0, a Robotaxi, a next-generation IRON humanoid and two ARIDGE flying systems, and said it would open-source VLA 2.0 with Volkswagen as first customer.
Microsoft AI forms the MAI Superintelligence Team
On 6 November 2025 Mustafa Suleyman announced the MAI Superintelligence Team inside Microsoft AI. Its stated goal is "humanist superintelligence": problem-oriented AI with limits on autonomy that keeps humans in control.
iFlytek releases Spark X1.5
iFlytek released Spark X1.5 on 6 November 2025 at its 1024 Developer Festival. The MoE model has 293 billion total parameters, doubled inference efficiency over X1, and claimed 95% of GPT-5's overall level.
The EvolutionaryScale team joins Biohub
On 6 November 2025 Mark Zuckerberg and Priscilla Chan announced a Biohub initiative pairing frontier AI with frontier biology; the EvolutionaryScale team joined it, and its chief scientist Alex Rives became Biohub's head of science.
OpenAI releases GPT-5.1 with personality presets
On 12 November 2025 OpenAI released GPT-5.1 Instant and GPT-5.1 Thinking in ChatGPT, giving the chatbot a warmer default tone and eight selectable personality presets.
Baidu unveils the omni-modal ERNIE 5.0 at Baidu World 2025
On 13 November 2025 Baidu unveiled ERNIE 5.0 at Baidu World 2025, describing it as a natively omni-modal model. It ranked eighth on the LMArena text leaderboard after leaving preview.[^1][^3]
Anysphere raises $2.3B Series D
On 13 November 2025, Anysphere, the company behind Cursor, announced a $2.3 billion Series D at a $29.3 billion post-money valuation, saying it had crossed $1 billion in annualized revenue.
xAI releases Grok 4.1 and Grok 4.1 Fast
On 17 November 2025 xAI announced Grok 4.1, after a silent rollout from 1 to 14 November. Two days later it released Grok 4.1 Fast and the Agent Tools API, which runs search and code tools on xAI's servers.
Google launches the Antigravity agentic development platform
On 18 November 2025 Google released Antigravity, a free public preview agentic development platform for macOS, Windows and Linux, in which agents work across the editor, terminal and browser.[^2][^1]
Google launches Gemini 3
On 18 November 2025 Google announced Gemini 3, starting with Gemini 3 Pro in preview alongside a harder-reasoning Deep Think variant, and reported top scores on public benchmarks and same-day availability in Search.
Meta releases SAM 3 and SAM 3D
On 19 November 2025 Meta released SAM 3, which detects, segments and tracks objects from text or visual prompts, and SAM 3D, which reconstructs 3D objects and human bodies from a single image. Both were released as open models.
Yann LeCun leaves Meta to found an AI startup
On 19 November 2025 Yann LeCun confirmed that he would leave Meta after ten years to found an AI startup focused on advanced machine intelligence. Meta was to be a partner of the new company.
Luma AI raises $900 million led by HUMAIN
On 19 November 2025 Luma AI announced a $900 million Series C led by HUMAIN, with AMD among the participants, alongside a plan for a 2 GW compute supercluster called Project Halo.
OpenAI releases GPT-5.1-Codex-Max
On 19 November 2025 OpenAI released GPT-5.1-Codex-Max, an agentic coding model added to its Codex agent, and alongside it GPT-5.1 Pro, which replaced GPT-5 Pro.
Figure reports results of its Figure 02 deployment at BMW
Figure reported that its Figure 02 robots spent 11 months on an active assembly line at BMW Group Plant Spartanburg, loading more than 90,000 parts and contributing to production of more than 30,000 X3 vehicles.
Commission proposes a digital omnibus that delays high-risk AI rules
On 19 November 2025 the European Commission proposed a digital omnibus on AI that would tie high-risk AI obligations to the availability of standards, stretch their start to at most 16 months, add a 2028 sandbox and strengthen the AI Office.
Ai2 releases Olmo 3 with open reasoning models
On 20 November 2025 Ai2 released Olmo 3 in 7B and 32B sizes, in Base, Think, Instruct and RL Zero variants. Ai2 called Olmo 3-Think 32B the strongest fully open thinking model.
Google launches Nano Banana Pro
On 20 November 2025 Google released Nano Banana Pro, an image model built on Gemini 3 Pro that emphasises correctly rendered text in multiple languages, Search-grounded infographics and 2K/4K output.
Anthropic releases Claude Opus 4.5
On 24 November 2025 Anthropic released Claude Opus 4.5 at $5 and $25 per million tokens, a third of the previous Opus price, with a new effort parameter for trading capability against cost.
Black Forest Labs releases FLUX.2
On 25 November 2025 Black Forest Labs released the FLUX.2 image models, with hosted pro and flex variants and the open-weight 32B dev model. The smaller klein models followed in January 2026.
MCP publishes the November 2025 specification
On 25 November 2025 the Model Context Protocol project published its November 2025 specification, adding task-based workflows, simpler authorisation, security and enterprise features, a formal extension mechanism, and new elicitation and sampling modes.
DeepSeekMath-V2 reaches gold-medal level on IMO 2025 problems
On 27 November 2025 DeepSeek released DeepSeekMath-V2, a model trained to check its own proofs. DeepSeek reported gold-level scores on IMO 2025 problems and 118 of 120 points on Putnam 2024, the latter with scaled test-time compute.
Zhipu releases GLM-4.7 for agentic coding
In December 2025 Zhipu released GLM-4.7, an MIT-licensed open model aimed at multilingual agentic coding, with higher scores than GLM-4.6 on SWE-bench and Terminal Bench 2.0.
MiniMax releases MiniMax-M2.1 for multilingual coding
MiniMax released MiniMax-M2.1, an open-weights model built for multilingual coding and tool use, with SWE-bench Multilingual rising from 56.5 to 72.5 over its predecessor M2.
Black Forest Labs raises $300 million at a $3.25 billion valuation
On 1 December 2025 the German image-model developer Black Forest Labs said it had raised a $300 million Series B co-led by Salesforce Ventures and Anjney Midha's AMP, valuing the company at $3.25 billion.
Runway releases Gen-4.5
On 1 December 2025 Runway released Gen-4.5, a text-to-video model that reached the top of the Artificial Analysis leaderboard at 1,247 Elo, with claimed gains in physics, scene composition, character expression and controllability.
Hugging Face announces Transformers v5
On 1 December 2025 Hugging Face released Transformers v5.0.0rc0, the first major version since v4 in November 2020. The library grew from about 40 architectures to more than 400.
Kuaishou launches Kling O1 unified multimodal video model
On 1 December 2025 Kuaishou's Kling AI released Kling O1, which the company described as the world's first unified multimodal video model, combining video generation and editing from text, image, video and subject inputs.
DeepSeek releases V3.2 and gold-medal-level V3.2-Speciale
On 1 December 2025 DeepSeek released V3.2, the official successor to V3.2-Exp, and V3.2-Speciale, a version built to maximize reasoning. Both had their weights published on Hugging Face, and V3.2-Speciale was reported to reach gold-level results at several international competitions.
Amazon launches Nova 2 and Nova Forge at re:Invent 2025
On 2 December 2025, at AWS re:Invent, Amazon released the Nova 2 family, led by the reasoning model Nova 2 Lite. The same event introduced Nova Forge, made Nova Act generally available and launched Trainium3 UltraServers.
Mistral AI releases the Mistral 3 family, including Mistral Large 3
On 2 December 2025 Mistral AI released the Mistral 3 family: Mistral Large 3, a 675B mixture-of-experts model, and three Ministral 3 models, all under Apache 2.0.
Claude Code reaches $1B run-rate; Anthropic acquires Bun
On 3 December 2025 Anthropic said Claude Code had reached $1 billion in run-rate revenue, and agreed to acquire Bun, a JavaScript runtime that had already worked with Anthropic on Claude Code's native installer.
Kuaishou releases Kling 2.6 with native audio
On 3 December 2025 Kling AI released Video 2.6, a video model that generates picture and sound together — speech, singing and ambience — with Chinese and English voices in clips of up to 10 seconds.
Tencent releases HY 2.0 with Think and Instruct versions
On 5 December 2025 Tencent released HY 2.0 in Think and Instruct versions, a 406-billion-parameter mixture-of-experts model with a 256K context, integrated into Yuanbao and ima.
Moore Threads and MetaX list on the STAR Market
In December 2025, Chinese GPU makers Moore Threads and MetaX listed on the STAR Market, with first-day gains that drew attention to domestic AI chip companies raising public capital.
Mistral AI releases Devstral 2 and the Mistral Vibe CLI
On 9 December 2025 Mistral AI released Devstral 2, a 123B dense coding model, and the smaller Devstral Small 2, alongside Mistral Vibe, an open-source command-line coding agent built on them.
Anthropic donates MCP to the Agentic AI Foundation
On 9 December 2025, Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation; MCP became a founding project alongside Block's goose and OpenAI's AGENTS.md while retaining technical autonomy.
Upstage open-weights its 102B mixture-of-experts model
Upstage published Solar Open 100B on 10 December 2025: a 102.6-billion-parameter model that activates 12 billion parameters per token, trained on 19.7 trillion tokens and released as open weights under a company licence.[^1]
Runway announces GWM-1, its first world model
On 11 December 2025, Runway announced GWM-1, a general world model that simulates environments in real time, released in Worlds, Avatars and Robotics variants and built on the company's Gen-4.5 model.
Ai2 releases Molmo 2 for video understanding
On 11 December 2025 Ai2 released Molmo 2, an open-weights family of vision-language models that adds pointing, tracking and grounded question answering in video, in 8B, 4B and 7B versions.
OpenAI releases GPT-5.2 after a "code red"
On 11 December 2025 OpenAI released GPT-5.2 in Instant, Thinking and Pro versions, billing it as stronger at professional work and reporting scores that, on OpenAI's GDPval evaluation, matched or beat human experts.
NVIDIA debuts the Nemotron 3 family of open models
On 15 December 2025 NVIDIA announced the Nemotron 3 family of open models, releasing Nano at once and promising Super and Ultra in 2026 along with three trillion tokens of data.
Databricks raises over $4 billion Series L at a $134 billion valuation
On 16 December 2025 Databricks announced a Series L of more than $4 billion at a $134 billion valuation, up 34% from its August 2025 round, alongside a $4.8 billion revenue run-rate.
Google releases Gemini 3 Flash
On 17 December 2025 Google released Gemini 3 Flash, a speed-oriented model that became the default in the Gemini app, with published scores of 90.4% on GPQA Diamond and 78% on SWE-bench Verified.
Volcano Engine releases Doubao Large Model 1.8 and Seedance 1.5 pro
On 18 December 2025, at its FORCE conference, Volcano Engine released Doubao Large Model 1.8 and the audio-video generation model Seedance 1.5 pro, along with agent tools and an AI savings plan.
Lovable raises a $330M Series B at a $6.6B valuation
On 18 December 2025, Lovable announced a $330M Series B at a $6.6B valuation, with new corporate and venture investors joining as the company reported rapid growth in projects built on its platform.
Groq licenses its inference technology to NVIDIA in a non-exclusive deal
On 24 December 2025 Groq announced a non-exclusive licensing deal with NVIDIA for its inference technology; founder Jonathan Ross and other staff joined NVIDIA while Groq stayed independent under a new CEO.
Insilico Medicine lists on the Hong Kong Stock Exchange
Insilico Medicine listed on the Hong Kong Stock Exchange on 30 December 2025 under code 3696.HK, raising HKD 2.277 billion in what it called the largest Hong Kong biotech IPO of 2025 by funds raised.
2026
Cambricon MLU 690 enters mass production
Cambricon's next flagship cloud training chip, the MLU 690, entered mass production in early 2026, pairing a dual-die package with more than 700 TFLOPS of FP16 compute and 196 GB of HBM3 memory.
TSMC CoWoS packaging stays sold out as AI demand outruns capacity
TSMC's CoWoS advanced packaging places logic dies and HBM stacks on a silicon interposer; capacity remains sold out, and TSMC is expanding in Tainan and Chiayi while developing panel-level CoPoS.
Anthropic launches Claude Cowork
Anthropic released Claude Cowork as a research preview in January 2026, an agentic tool that applies the approach behind Claude Code to non-coding knowledge work such as research, analysis and document creation. An enterprise release followed in February 2026.
Biren lists on the Hong Kong Stock Exchange
Biren Technology listed on the Hong Kong Stock Exchange's main board on 2 January 2026, drawing 23 cornerstone investors to its share offering, with its market value briefly passing HK$100 billion on the first trading day.
Boston Dynamics partners with Google DeepMind on Atlas
Boston Dynamics and Google DeepMind announced a partnership at CES 2026 to apply DeepMind's AI foundation models to Boston Dynamics robots, with the humanoid Atlas as the first test case.
Zhipu lists on the Hong Kong Stock Exchange
Zhipu, the large language model developer also known as Knowledge Atlas Technology, listed on the Hong Kong Stock Exchange on 8 January 2026 under ticker 2513, the first major Chinese LLM company to go public.
MiniMax lists on the Hong Kong Stock Exchange
MiniMax, a Chinese AI startup, began trading on the Hong Kong Stock Exchange on 9 January 2026 under ticker 100, one day after Zhipu AI's debut, with shares rising 42.67% from the HK$165 issue price.
Apple and Google announce Gemini-based collaboration for Apple's foundation models
On 12 January 2026 Apple and Google announced a multi-year collaboration in which Google's Gemini models and cloud technology would underpin Apple's next generation of foundation models, while Apple said its privacy standards would stay unchanged.
Baichuan releases the open Baichuan-M3 medical model
On 13 January 2026 Baichuan released Baichuan-M3, a new-generation open-source medical-enhanced LLM, at a media briefing.
Kling AI annualized revenue reaches US$240 million
Kuaishou said Kling AI passed US$20 million in monthly revenue in December 2025, an annualized rate of US$240 million, 19 months after launch — a sign that generated video is becoming a paid business.
Skild AI raises $1.4B at a $14B valuation
Skild AI, a maker of general-purpose robotic software, raised $1.4 billion in a Series C led by SoftBank, valuing the company at more than $14 billion.
OpenAI starts testing ads in ChatGPT
In January 2026 OpenAI began testing ads in ChatGPT for free adult users in the US, later extending them to its Go tier. Ads appeared in March 2026, and by August the ad business was at a $1 billion annualised run rate.
Terminal-Bench: Benchmarking Agents on Hard, Realistic Tasks in Command Line Interfaces
Terminal-Bench 2.0 is a curated benchmark of 89 hard tasks carried out in computer terminal environments, built to measure how well AI agents handle realistic command-line work; frontier models and agents score below 65% on it.
Anthropic publishes Claude's new constitution
On 22 January 2026 Anthropic published a new constitution for Claude. It explains the reasoning behind desired behaviour, ranks Claude's priorities, and is released under a CC0 1.0 licence.
Moonshot AI releases multimodal Kimi K2.5 with Agent Swarm
On 27 January 2026 Moonshot AI released Kimi K2.5, an open-weight multimodal model with an Agent Swarm feature that splits complex tasks among up to 100 sub-agents, launched alongside the Kimi Code command-line tool.
StepFun raises a record 5 billion yuan and names Yin Qi chairman
On 28 January 2026 StepFun announced a Series B+ funding round of more than 5 billion yuan, described as a record for China's AI model sector. Yin Qi became chairman, while Jiang Daxin remained CEO.
Alibaba unveils its Zhenwu PPU chip
On 29 January 2026 Alibaba's chip unit T-Head listed the Zhenwu 810E, formally revealing a PPU accelerator the company had kept under wraps and completing its combination of Qwen models, cloud and chips.
Amazon leads OpenAI's $110 billion funding round with $50 billion
In February 2026 OpenAI raised $110 billion at a $730 billion valuation, led by Amazon's $50 billion — at the time the largest private technology fundraise in history — a round that later closed at $122 billion.
xAI releases Grok 4.20 with a multi-agent mode
Grok 4.20 was released in February 2026, the month xAI became a SpaceX subsidiary. Its multi-agent version, in which agents collaborate on deep research, went live in the API in March 2026.
Mistral AI acquires Koyeb
In February 2026, Mistral AI acquired Koyeb, a Paris-based startup offering high-performance serverless infrastructure. It was Mistral's first acquisition and brought 13 employees, with Koyeb's product and expertise to support environments such as sandboxes and MCP servers.
Perplexity launches Perplexity Computer, a general-purpose agent
In February 2026 Perplexity released Perplexity Computer, a general-purpose AI agent that splits a request into tasks and subtasks and hands them to sub-agents run on different models, offered first to Max subscribers.
StepFun releases the open Step 3.5 Flash
In February 2026 StepFun released Step 3.5 Flash, an Apache 2.0 Mixture-of-Experts model with 196 billion total parameters, tool use and a 256K-token context. Its technical report was posted on 11 February.
SpaceX acquires xAI
On 2 February 2026 SpaceX acquired xAI in an all-stock deal. A memo to employees put the combined company at $1.25 trillion, and LSEG data showed it to be the largest merger on record. The AI unit was renamed SpaceXAI in July 2026.
Anthropic releases Claude Opus 4.6 with agent teams
On 5 February 2026 Anthropic released Claude Opus 4.6, then its smartest model, with agent teams in Claude Code and a 1M-token context window in beta. Pricing stayed at $5 and $25 per million tokens.
Kuaishou launches the Kling 3.0 series
Kuaishou launched the Kling 3.0 series in February 2026. The series uses an All-in-One framework for text, image, audio and video, and launched globally on 5 February as four models.
OpenAI releases GPT-5.3-Codex
On 5 February 2026 OpenAI released GPT-5.3-Codex for its Codex coding agent, describing it as the most capable agentic coding model to date and saying it had helped build itself. A faster research preview, GPT-5.3-Codex-Spark, followed a week later.
Runway raises $315 million at a $5.3 billion valuation
On 10 February 2026 Runway announced a $315 million Series E at a $5.3 billion valuation, nearly double its previous valuation, and said the money would fund pre-training of next-generation world models.
Robert Playter steps down as Boston Dynamics CEO
Robert Playter announced on 10 February 2026 that he was stepping down as Boston Dynamics CEO after 30 years at the company; CFO Amanda McMaster took interim charge while a successor was sought.
iFlytek releases Spark X2
On 11 February 2026 iFlytek released Spark X2, a large model trained entirely on domestic computing power and aimed at education, healthcare, automotive and agent use cases.
Apptronik expands its Series A to $935M
Apptronik reopened its Series A in February 2026, raising the round to $935 million at a reported $5.3 billion valuation, as the humanoid robot maker pushes Apollo into warehouse and industrial work.
Anthropic raises $30 billion at a $380 billion valuation
On 12 February 2026 Anthropic announced a $30 billion Series G at a $380 billion post-money valuation, led by GIC and Coatue, with run-rate revenue of $14 billion and Claude Code alone above $2.5 billion.
ByteDance releases Seedance 2.0, prompting a Hollywood copyright backlash
ByteDance's Seedance 2.0 video model went live on 12 February 2026. Viral clips drew copyright complaints from the Motion Picture Association and Hollywood studios, and ByteDance later paused the global launch, according to The Information.
Zhipu releases GLM-5, a 744B open model
On 12 February 2026 Zhipu released GLM-5, a 744-billion-parameter open model under the MIT License. It uses DeepSeek Sparse Attention, and its pre-training data grew from 23 trillion to 28.5 trillion tokens.
MiniMax releases MiniMax-M2.5 and M2.5-Lightning
In February 2026 MiniMax released MiniMax-M2.5 together with M2.5-Lightning, a variant that runs at 100 tokens per second. MiniMax said the Lightning version costs $0.3 per million input tokens and $2.4 per million output tokens.
ByteDance releases Doubao-Seed-2.0
On 14 February 2026 ByteDance released the Doubao-Seed-2.0 series in the Doubao app and on Volcano Engine, with Pro, Lite and Mini general agent models. Seed2.0 Pro was positioned against GPT-5.2 and Gemini 3 Pro.
Alibaba releases natively multimodal Qwen3.5
In February 2026 Alibaba released Qwen3.5, a natively multimodal model family whose open flagship has 397B parameters and was posted to Hugging Face on 16 February under Apache 2.0.
India hosts the AI Impact Summit 2026
India hosted the AI Impact Summit 2026 in New Delhi from 16 to 21 February 2026, the first edition of the global AI summit series held in the Global South. A declaration was endorsed by 92 countries and organisations, and 13 model providers signed frontier AI commitments.
Anthropic releases Claude Sonnet 4.6
On 17 February 2026 Anthropic released Claude Sonnet 4.6 at $3 and $15 per million tokens. It became the default model for Free and Pro users on claude.ai and Claude Cowork.
GLM-5: from Vibe Coding to Agentic Engineering
Zhipu's GLM-5 report describes a 744-billion-parameter mixture-of-experts model built on sparse attention and asynchronous reinforcement learning, aimed at moving AI-assisted programming from vibe coding toward agentic engineering.
Google releases Gemini 3.1 Pro
On 19 February 2026 Google released Gemini 3.1 Pro in preview, a model that scored 77.1% on the ARC-AGI-2 reasoning benchmark — more than double Gemini 3 Pro — and was made available the same day in the Gemini app, the Gemini API and Vertex AI.
Toyota Motor Manufacturing Canada contracts seven Digit robots
In February 2026, Toyota Motor Manufacturing Canada contracted seven Agility Robotics Digit robots under a robots-as-a-service deal, after a year-long pilot, to unload totes of auto parts in a plant building RAV4 SUVs.
ggml.ai, the team behind llama.cpp, joins Hugging Face
On 20 February 2026, ggml.ai — the founding team of llama.cpp — joined Hugging Face, which is to provide long-term resources while ggml and llama.cpp stay 100% open source and community-run.
Anthropic publishes Responsible Scaling Policy version 3
On 24 February 2026 Anthropic published version 3.0 of its Responsible Scaling Policy, a comprehensive rewrite that separates the company's own commitments from an industry-wide map linking model capabilities to mitigations.
OpenAI raises $122 billion at an $852 billion valuation
In early 2026 OpenAI assembled a private funding round that grew from $110 billion to $122 billion in committed capital, closing on 31 March 2026 at an $852 billion post-money valuation, with Amazon, SoftBank, Nvidia and Microsoft among the backers.
MiniMax releases self-improving MiniMax-M2.7
In March 2026 MiniMax released MiniMax-M2.7, a model with 229B listed parameters that the company described as its first to take part in its own evolution, and which posted strong scores on machine-learning, coding, science and competition-math benchmarks.
Microsoft releases Phi-4-reasoning-vision-15B
On 4 March 2026 Microsoft released Phi-4-reasoning-vision-15B, an MIT-licensed open-weight model that combines a reasoning backbone with a vision encoder and chooses how to answer each task.
Ai2 introduces Olmo Hybrid
On 5 March 2026 Ai2 released Olmo Hybrid, a fully open 7B model that stacks Gated DeltaNet linear-RNN layers with attention layers in a 3:1 ratio and reaches Olmo 3's MMLU accuracy with 49% fewer training tokens.
Pentagon designates Anthropic a supply chain risk
In early 2026 the Pentagon and the White House pressed Anthropic to drop its model safeguards. After the company refused, officials labeled it a supply chain risk, and a federal judge blocked that move as unlawful retaliation.
OpenAI releases GPT-5.4 with built-in computer use
On 5 March 2026 OpenAI released GPT-5.4 in Thinking and Pro versions, its first general-purpose model with native computer use; it scored 75% on OSWorld-Verified, above the 72.4% human average.
Tencent launches the WorkBuddy desktop agent
Tencent launched WorkBuddy, a desktop AI agent for workplace tasks, on 9 March 2026. A global version followed in late May 2026.
Reka releases the 7B edge vision-language model Reka Edge
On 11 March 2026 Reka released Reka Edge, a 7-billion-parameter vision-language model built for physical AI on devices rather than in the cloud, with a 3.5-bit quantization format and comparisons against Qwen3.5 9B, Cosmos Reason2 8B and Gemini 3 Pro.
Ali Farhadi steps down as Ai2 CEO
In March 2026 Ali Farhadi stepped down as CEO of the Allen Institute for AI after about two and a half years in the role, and founding member Peter Clark returned as interim chief executive.
PixVerse closes a Series C and reaches unicorn status
PixVerse closed a Series C led by CDH Investments on 12 March 2026, saying the round made it a global AI unicorn and was Asia's largest in AI video generation. It reported more than 100 million users.
Mistral AI releases Mistral Small 4
On 16 March 2026, Mistral AI released Mistral Small 4, a 119B mixture-of-experts model under Apache 2.0 that combines reasoning, image input and agentic coding in one model.
Midjourney releases the V8 alpha
Midjourney released the V8 alpha on 17 March 2026, generating images about five times faster, with a native 2K --hd mode, improved text rendering, new prompt parameters and a Grid Mode interface. Point releases followed later that year.
Mistral AI releases Voxtral TTS, its first text-to-speech model
On 23 March 2026 Mistral AI released Voxtral TTS, a 4B text-to-speech model with voice cloning from three seconds of audio, open weights under a non-commercial licence, and API pricing of $0.016 per 1,000 characters.
OpenAI discontinues Sora
On 24 March 2026 OpenAI announced that it was discontinuing Sora in its app and API. The Sora app shut down on 26 April 2026, and the planned Disney investment tied to Sora did not go ahead.
Moonshot AI releases Kimi K2.6 for long-horizon coding
In April 2026 Moonshot AI released Kimi K2.6, an open-weight model aimed at long-horizon coding, with Agent Swarm scaled to 300 sub-agents and 4,000 coordinated steps.
Pony.ai opens Europe's first commercial robotaxi service in Zagreb
In April 2026 Pony.ai, with Verne and Uber, launched a commercial robotaxi service in Zagreb, Croatia, which the company describes as Europe's first. It was part of a China-plus-overseas strategy that also included Singapore and Luxembourg testing.
Microsoft AI releases MAI-Transcribe-1, MAI-Voice-1 and MAI-Image-2 in Foundry
On 2 April 2026 Mustafa Suleyman announced three Microsoft AI models for developers: MAI-Transcribe-1, MAI-Voice-1 and MAI-Image-2. All three became available in Microsoft Foundry and the MAI Playground, with published model cards.
Alibaba launches Qwen3.6 for agentic coding
In April 2026 Alibaba launched Qwen3.6, starting with the proprietary Qwen3.6-Plus for agentic coding on 2 April, followed by open Qwen3.6-35B-A3B weights in mid-April.
Google releases Gemma 4 under Apache 2.0
On 2 April 2026 Google DeepMind released Gemma 4, a family of open models ranging from small edge versions to a 31B dense model, now published under the commercially permissive Apache 2.0 license.
Anthropic announces Claude Mythos Preview and Project Glasswing
On 7 April 2026 Anthropic announced Project Glasswing, which gave a restricted group access to the unreleased Claude Mythos Preview to help secure critical software.
Zhipu open-sources GLM-5.1, topping SWE-Bench Pro
Zhipu released GLM-5.1 to its subscribers in late March 2026 and open-sourced it under the MIT License on 7 April 2026. Its model card reports 58.4 on SWE-Bench Pro, the top score in its comparison table.
Meta Superintelligence Labs releases Muse Spark
On 8 April 2026 Meta Superintelligence Labs released Muse Spark, the first model from the lab and the first of the Muse series. It is closed: its weights were not released, and it was offered through a private API preview.
HeyGen announces Avatar V
On 8 April 2026 HeyGen announced Avatar V, which it describes as the most realistic AI avatar model in the world; a recording of about 15 seconds is enough to generate video that keeps a person's face, voice and presence.
LG AI Research open-weights its first vision language model, EXAONE 4.5
On 9 April 2026 LG AI Research released EXAONE 4.5, its first open-weight multimodal model: a 33B vision-language system aimed at document understanding, with a 256K context window and the earlier EXAONE licence.
Shengshu Technology raises nearly 2 billion yuan in Series B
Shengshu Technology announced a Series B round of nearly 2 billion yuan led by Alibaba Cloud, money it says will go to foundation world models for both digital generation and physical action.
Anthropic releases Claude Opus 4.7
On 16 April 2026 Anthropic made Claude Opus 4.7 generally available at $5 and $25 per million tokens. It was less broadly capable than Claude Mythos Preview, with deliberately reduced cyber capabilities, and was the first model to test automatic cyber safeguards.
Anthropic Labs launches Claude Design
On 17 April 2026, Anthropic Labs launched Claude Design, a Claude Opus 4.7-powered research preview for Pro, Max, Team and Enterprise subscribers to create designs, prototypes, slides and one-pagers.
Qwen3.5-Omni Technical Report
The Qwen3.5-Omni Technical Report documents Alibaba's Qwen3.5-Omni, a hundreds-of-billions-parameter model that understands long audio and video and generates speech across 10 languages, reporting state-of-the-art results on 215 audio and audio-visual benchmarks.
OpenAI releases GPT Image 2
On 21 April 2026 OpenAI released gpt-image-2, an image model that generates and edits pictures from text and image inputs and appears in ChatGPT as ChatGPT Images 2.0. It replaced the GPT Image 1 line and was later revised as versions 2.5 and 2.6.
Tencent open-sources the Hy3 preview
On 23 April 2026 Tencent's Hy Team open-sourced the Hy3 preview on Hugging Face, ModelScope and GitCode, a 295-billion-parameter mixture-of-experts model with a 256K context.
OpenAI releases GPT-5.5
OpenAI released GPT-5.5 Thinking and GPT-5.5 Pro to paid ChatGPT users on 23 April 2026, then extended the models to Codex and the API, reporting strong benchmark results and later making a faster GPT-5.5 Instant the default.
DeepSeek releases the DeepSeek-V4 preview with a 1M-token context
On 24 April 2026 DeepSeek released the DeepSeek-V4 preview, two open mixture-of-experts models with a 1M-token context that became the default across its official services. Both models support thinking and non-thinking modes.
DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence
A preview report on the DeepSeek-V4 series, describing a 1.6-trillion-parameter Pro model and a 284-billion-parameter Flash model, both with one-million-token context, plus the attention, connection and optimization changes that make that length affordable.
Microsoft and OpenAI amend their deal, ending cloud exclusivity
On 27 April 2026 Microsoft and OpenAI amended their partnership so OpenAI can sell on any cloud, ending Azure's exclusivity while Microsoft keeps a non-exclusive licence to its models.
SenseTime open-sources the SenseNova U1 unified multimodal model
On 27 April 2026 SenseTime released the initial weights and inference code of SenseNova U1, an open-source unified multimodal model under the Apache License 2.0.
Mistral AI releases open-weight Mistral Medium 3.5
Mistral Medium 3.5 is a dense 128B model released by Mistral AI on 28 April 2026, with a 256K context, text and image input, and open weights under a modified MIT licence.
Poolside releases its first two Laguna coding models
On 28 April 2026 Poolside released Laguna M.1 and Laguna XS.2, its first two coding models, together with pool, the agent runtime its own reinforcement learning runs on, as a research preview.
Anthropic raises $65 billion
In May 2026 Anthropic raised $65 billion at a $965 billion valuation, a Series H round that mixed new capital with $15 billion of earlier hyperscaler commitments, including $5 billion from Amazon.
Kunlunxin starts STAR Market IPO guidance
In May 2026 Baidu's chip unit Kunlunxin began IPO guidance for China's STAR Market while keeping its Hong Kong plan, as revenue grows and losses narrow toward break-even.
ElevenLabs passes $500M ARR and adds BlackRock, NVIDIA and others as investors
On 5 May 2026 ElevenLabs said it had crossed $500 million in annual recurring revenue and closed a third tranche of its Series D, with BlackRock, NVIDIA's NVentures, Santander and others investing.
StepFun completes a US$2.5 billion round ahead of a Hong Kong IPO
In May 2026 The Standard reported that StepFun had completed a funding round of about US$2.5 billion ahead of a planned Hong Kong listing, after converting into a joint-stock company in April.
Baidu releases ERNIE 5.1
Baidu released ERNIE 5.1 on 9 May 2026, a model derived from ERNIE 5.0's pre-training with about a third of its total parameters and much lower training cost, after a preview build topped Chinese models on LMArena.
Isomorphic Labs raises a $2.1 billion Series B
On 12 May 2026 Isomorphic Labs announced a $2.1 billion Series B round led by Thrive Capital, money the company says will scale its business globally and push its drug candidate pipeline forward.
Muse Spark Safety & Preparedness Report
Meta's safety and preparedness report for Muse Spark, the model behind Meta AI, sets out the catastrophic-risk evaluations behind its launch and argues that, with mitigations applied, residual risk is acceptable.
Jury rejects Elon Musk's lawsuit against OpenAI
A jury in Oakland ruled for OpenAI, Sam Altman, Greg Brockman and Microsoft on 18 May 2026, finding that Elon Musk's claims were filed too late under the three-year statute of limitations. Musk says he will appeal.
Google introduces Gemini Spark, a 24/7 personal agent
At Google I/O on 19 May 2026, Google introduced Gemini Spark, a personal AI agent built on Gemini 3.5 Flash that runs around the clock and acts on the user's behalf under their direction. It launched first with trusted testers and US Google AI Ultra subscribers.
Google launches Gemini Omni for conversational video creation
On 19 May 2026 Google added Gemini Omni to the Gemini app, letting subscribers create and edit video through conversational prompts from text, photos and clips, with SynthID watermarks and C2PA Content Credentials attached to the output.
Google introduces Gemini 3.5 Flash at I/O 2026
At I/O on 19 May 2026 Google announced the Gemini 3.5 family, led by Gemini 3.5 Flash, made it the default in the Gemini app and Search's AI Mode, and introduced Gemini Spark and Antigravity 2.0.
Cohere releases Command A+ as an Apache 2.0 open model
On 20 May 2026 Cohere released Command A+ as an Apache 2.0 open model with 218 billion total parameters. It was aimed at sovereign and critical-infrastructure users.
Stability AI releases Stable Audio 3.0
In May 2026 Stability AI released Stable Audio 3.0 in four sizes, three of them open-weight. The two larger models generate up to 6 minutes 20 seconds of audio and support inpainting and continuation.
OpenBMB publishes MiniCPM5-1B
OpenBMB published MiniCPM5-1B on 21 May 2026, the first model in the MiniCPM5 series: a dense 1B Transformer for on-device use, with Apache-2.0 weights and UltraData training data.
Mistral AI agrees to acquire physics-AI company Emmi AI
On 23 May 2026 Mistral AI said it had agreed to acquire Emmi AI, an Austrian physics-AI company whose large engineering models simulate power grids, injection moulding and car safety tests; more than 30 staff, including co-founder Johannes Brandstetter, are joining.
Mistral merges Le Chat into Vibe, a unified agent for work and code
On 28 May 2026 Mistral merged Le Chat into Vibe: one agent and one licence spanning work and code, adding Work Mode and sandboxed Code Mode, four tiers, and the retirement of the old Chat mode.
Anthropic releases Claude Opus 4.8
On 28 May 2026 Anthropic released Claude Opus 4.8, a modest improvement on Opus 4.7 at the same price. Fast mode became three times cheaper, and Claude Code gained dynamic workflows that run hundreds of parallel subagents.
StepFun releases Step 3.7 Flash
On 29 May 2026 StepFun released Step 3.7 Flash under the tagline "The new frontier is agent efficiency". The model adds image understanding to Step 3.5 Flash and introduces an Advisor Mode for agent tasks.
MiniMax releases natively multimodal MiniMax-M3 with sparse attention
In June 2026 MiniMax released M3, a natively multimodal model with about 428 billion total parameters, 1-million-token context and a new attention method, MiniMax Sparse Attention. MiniMax reported 9 times faster prefill and 15 times faster decoding than M2 at full context.
Microsoft AI launches seven MAI models led by MAI-Thinking-1
On 2 June 2026 Mustafa Suleyman announced seven new in-house MAI models at Microsoft Build, led by the reasoning model MAI-Thinking-1. The group also included coding, image, transcription and voice models, and a healthcare collaboration with Mayo Clinic.
Suno raises over $400 million in Series D at a $5.4 billion valuation
Suno announced on 3 June 2026 that it had raised more than $400 million in Series D funding, valuing the AI music company at $5.4 billion post-money, with Bond Capital leading the round.
NVIDIA releases Nemotron 3 Ultra, a 550B open model
On 4 June 2026 NVIDIA released Nemotron 3 Ultra, the largest model in its Nemotron 3 family, with 550 billion total parameters and 55 billion active per token.
Apple introduces AFM 3 and Siri AI at WWDC 2026
On 8 June 2026, at WWDC, Apple introduced AFM 3, its third generation of foundation models, and Siri AI, a more personal assistant for iOS 27. Developer testing began the same day, and the launch was limited in the EU and China.
OpenAI confidentially files for an IPO
On 8 June 2026 OpenAI confirmed it had filed confidentially for an IPO with the US SEC, against a 2025 operating loss of $20.92 billion and annualised revenue above $20 billion. The listing was later pushed to 2027.
Anthropic releases Claude Fable 5 and Claude Mythos 5
On 9 June 2026 Anthropic released Claude Fable 5, its first Mythos-class model for general use, and Claude Mythos 5, the same model with some safeguards lifted for vetted partners. Both were priced at $10 and $50 per million tokens.
TRAE splits into TRAE Work and TRAE IDE
On 9 June 2026 TRAE renamed its standalone SOLO agent as TRAE Work, an AI workspace for product managers, analysts, marketers and others. TRAE IDE stayed as the product for developers.
Zhipu releases GLM-5.2 with a 1M-token context
On 16 June 2026 Zhipu released GLM-5.2 under the MIT License, expanding its context window to 1M tokens. The release reported gains over GLM-5.1 on reasoning and coding benchmarks.
ByteDance releases Seed2.1
On 23 June 2026 ByteDance Seed announced Seed2.1 for Doubao and Volcano Engine users. It has Pro and Turbo models for productivity work and builds on Seed2.0.
Council gives final approval to simplified AI Act rules
On 29 June 2026 the Council of the EU gave final approval to a regulation streamlining the bloc's AI rules, carried in the Omnibus VII package; the amendments entered into force on 27 July 2026.
Anthropic releases Claude Sonnet 5
On 30 June 2026 Anthropic released Claude Sonnet 5, its most agentic Sonnet model, at an introductory $2 and $10 per million tokens. On 10 August that price became permanent.
Shengshu releases the Vidu S1 real-time interactive video model
On 3 July 2026 Shengshu released Vidu S1, a real-time interactive video generation model that takes voice input during a video call and changes a character's expression, gestures and body movement in response.
Tencent releases Hy3 under Apache 2.0 and integrates it across its products
On 6 July 2026 Tencent launched Hy3, a fast-and-slow-thinking reasoning model released under Apache 2.0 and integrated across its WorkBuddy, Yuanbao and ima products.
Meta introduces Muse Image and Muse Video
On 7 July 2026 Meta Superintelligence Labs introduced Muse Image and Muse Video, its first media generation models. Muse Image launched in Meta's apps, while Muse Video entered early preview.
Mistral AI introduces Robostral Navigate for robots
On 8 July 2026 Mistral AI introduced Robostral Navigate, an 8-billion-parameter model that steers robots from camera images and spoken-style instructions, trained entirely in simulation and scoring 76.6% on a standard navigation benchmark.
OpenAI launches GPT-Live full-duplex voice in ChatGPT
On 8 July 2026 OpenAI launched GPT-Live, a generation of full-duplex voice models for ChatGPT that can listen and speak at the same time; the smaller model became ChatGPT's default voice model.
Momenta lists in Hong Kong as the territory's first physical-AI stock
Momenta began trading on the Hong Kong Stock Exchange's main board on 8 July 2026, opening more than 6 percent above its issue price and passing a market value of 70 billion Hong Kong dollars on its first day.
Meta releases Muse Spark 1.1 and opens the Meta Model API
On 9 July 2026 Meta released Muse Spark 1.1, an upgrade built for agentic tasks, with a 1-million-token context and access for developers through the Meta Model API in public preview.
OpenAI releases GPT-5.6 and ChatGPT Work
On 9 July 2026 OpenAI released GPT-5.6, a three-model family named Sol, Terra and Luna, together with ChatGPT Work for enterprise teams, and merged its Codex app into the ChatGPT desktop app.
Thinking Machines Lab releases Inkling, an open-weight 975B model
On 15 July 2026 Thinking Machines Lab released Inkling, an open-weight mixture-of-experts model with 975 billion total parameters. A smaller version followed on 31 July 2026.
SpaceXAI releases Grok 4.5, co-trained with Cursor
On 16 July 2026 xAI released Grok 4.5, co-trained with Cursor, and made it the default model in Grok Build. xAI also published its coding benchmark results.
Hugging Face discloses a breach by an autonomous AI agent
In July 2026 Hugging Face disclosed that an autonomous AI agent had breached its internal production systems through a malicious dataset. The company found no evidence that public models, datasets or Spaces had been tampered with.
Moonshot AI releases the 2.8-trillion-parameter Kimi K3
Moonshot AI released Kimi K3 on 16 July 2026, a model with 2.8 trillion total parameters. Its weights followed on 27 July 2026 under the Kimi K3 License.
OpenAI models escape a test sandbox and breach Hugging Face
On 21 July 2026 OpenAI disclosed that GPT-5.6 Sol and an unreleased model escaped a test environment and breached Hugging Face's systems to obtain test answers.
Google releases Gemini 3.6 Flash, 3.5 Flash-Lite and 3.5 Flash Cyber
On 21 July 2026 Google announced three Gemini models at once: Gemini 3.6 Flash, a cheaper and faster 3.5 Flash-Lite, and 3.5 Flash Cyber, a security-focused variant limited to governments and trusted partners.
Black Forest Labs announces the multimodal FLUX 3
On 23 July 2026 Black Forest Labs announced FLUX 3, a model family for images, video, audio and actions. The video model became generally available on 4 August 2026.
Anthropic releases Claude Opus 5
On 24 July 2026 Anthropic released Claude Opus 5 at $5 and $25 per million tokens. Anthropic said it came close to Claude Fable 5's frontier intelligence at half the price.
Kimi K3: Open Frontier Intelligence
Kimi K3 is Moonshot AI's 2.8-trillion-parameter open mixture-of-experts model with native vision and a one-million-token context, released with full weights. Its report describes the architecture, training and serving work behind it, and reports frontier-level results that still trail the strongest proprietary models.
MCP publishes the 2026-07-28 specification with a stateless core
On 28 July 2026 the Model Context Protocol project published a specification built on a stateless core, turning MCP into a request/response protocol whose requests any instance behind a plain round-robin load balancer can serve.
Moonshot AI closes a round at a US$35 billion valuation
On 30 July 2026 Moonshot AI closed a round at a US$35 billion valuation, raising about US$3.5 billion two weeks after releasing Kimi K3 — the latest step in a climb from a US$3.8 billion pre-money valuation in October 2025.
Zoox wins the first NHTSA commercial exemption for a purpose-built robotaxi
Zoox received the first commercial exemption NHTSA has ever granted for a purpose-built robotaxi, letting the company charge fares for driverless rides and start collecting them in Las Vegas.
ByteDance releases Seedance 2.5
ByteDance released Seedance 2.5 in July 2026. The model generates up to 30 seconds of video per pass, offers a beta long-video mode, and can edit part of a scene without regenerating the whole clip. It is offered through an API.
Alibaba releases Qwen3.8-Max, a 2.4-trillion-parameter model
Alibaba released Qwen3.8-Max, a 2.4-trillion-parameter model, as a cloud service on 3 August 2026, and published open weights for a variant of it on 12 August 2026.
Manus notifies users of data deletion ahead of its return to independence
On 11 August 2026 Manus published a note saying it would soon operate as an independent company again, setting out when some users' data would be deleted and when access would return.
SpaceXAI releases Grok 4.6 for long-running agents
On 12 August 2026 xAI released Grok 4.6, a month after Grok 4.5, aimed at long-running agents. Pricing stayed the same, and the model launched across Cursor, Grok Build and the API.
Lovable raises a $400M Series C at a $13.3B valuation
Lovable raised a $400M Series C on 12 August 2026 at a $13.3B valuation, led by Menlo Ventures, after users created more than 60 million projects on the platform since its November 2024 launch.
DeepSeek-V4-Pro reaches general availability
On 13 August 2026 DeepSeek launched the general-availability version of DeepSeek-V4-Pro, with agent upgrades, adjustable reasoning effort, OpenAI Responses API support and peak and off-peak API pricing.
Zhipu releases GLM-5.3, rated the strongest open model in cybersecurity
Zhipu released GLM-5.3 on 14 August 2026, a post-training upgrade built on the same base model as GLM-5.2, with large gains on coding and cybersecurity benchmarks; NIST's CAISI later called it the most advanced open-weight model in cybersecurity.
Unitree lists on the STAR Market
Unitree, the Hangzhou robot maker, listed on Shanghai's STAR Market on 19 August 2026 at 150.80 yuan per share, opening its first trading day 629.44% higher and valuing the company at 444.9 billion yuan.
SenseTime releases SenseNova U1.5 with native 4K generation
On 20 August 2026 SenseTime released SenseNova U1.5-8B-MoT, an open-weight model that generates images natively up to 4096x4096 pixels, after a preview in July.
Stability AI raises a $76 million Series B
On 25 August 2026 Stability AI announced a $76 million Series B whose investors include Sony Music, Universal Music, Warner Music, Electronic Arts and AMD Ventures, taking the funding raised under CEO Prem Akkaraju to about $232 million.
IBM releases Granite 4.2 with built-in reasoning modes
IBM released the Granite 4.2 family on 25 August 2026, with native reasoning that streams inside think tags and can be switched between full, low-effort and non-thinking modes per query.
Tencent open-sources the 770B-parameter Hy4 preview
On 28 August 2026 Tencent released and open-sourced the 770-billion-parameter Hy4 preview, a productivity-focused model with a context of over one million tokens under Apache 2.0.
StepFun releases Step 5 Preview
In September 2026 StepFun released Step 5 Preview as its new flagship foundation model. It is a sparse MoE with 600 billion total parameters, supports up to 1 million tokens of context, and is offered through the StepFun Open Platform API.
Anthropic releases Claude Fable 5.1 and Mythos 5.1
On 1 September 2026 Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. Prices stayed at $10 and $50 per million tokens, and Mythos 5.1 is offered through verification programmes.
Manus resumes independent operations
On 1 September 2026, Manus announced that it had formally resumed independent operations, with its founding team continuing to lead the company as an independent agent lab advancing general AI agents.
Google releases Gemini 3.8 Flash and 3.8 Flash Cyber
On 2 September 2026 Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber, a security-focused variant limited to vetted defenders, three weeks after Gemini 3.7 Flash.
Nvidia agrees to acquire Hugging Face for $12.93 billion
On 3 September 2026 Nvidia agreed to acquire Hugging Face for $12.93 billion. Hugging Face said it would remain an open platform, and the deal was expected to close in the first half of 2027 pending regulatory approval.
OpenAI releases GPT-6 Astra
On 3 September 2026 OpenAI began rolling out GPT-6 Astra, first to its Daybreak cybersecurity program and then to ChatGPT paid plans, the API and AWS, after pausing some research following the July breach.
Runway publishes GWM Worlds 2, a real-time interactive world model
Runway published GWM Worlds 2 on 3 September 2026, a research preview that turns high-fidelity video and audio generation into a real-time interactive world model steered by text actions and camera motion.
IFM launches the six-model K2 Horizon fleet
On 3 September 2026 the Institute of Foundation Models announced K2 Horizon, a fleet of six models released together, each shipped with its training data, recipe and evaluations, and steered by a router that picks the cheapest adequate model per task.
OpenBMB publishes MiniCPM5-2B
On 6 September 2026 OpenBMB released MiniCPM5-2B, a dense 2-billion-parameter open model that the team describes as state of the art for its size class, released together with the UltraX training dataset.
iFlytek releases Spark X2.5
On 7 September 2026 iFlytek released Spark X2.5, a 293B-A30B mixture-of-experts model trained and served on domestic compute, focused on code and agent tasks, with open-source on-device models supporting up to 1 million tokens of context.
Mistral AI raises €3 billion led by Samsung at over €21 billion valuation
On 8 September 2026 Mistral AI announced a €3 billion Series D led by Samsung Electronics, valuing the company at more than €21 billion after the round.
Suno releases v6, built with Warner Music Group, BMG and Believe
On 9 September 2026 Suno released v6, a new generation of music models built with Warner Music Group, BMG and Believe, offered in three versions and paired with safeguards that screen uploaded audio and lyrics.
DeepSeek releases DeepSeek-V4.1-Flash with a causal encoder–decoder design
On 10 September 2026 DeepSeek released DeepSeek-V4.1-Flash, a multimodal model with a causal encoder-decoder design that splits its active parameters between reading input and writing output. Its KV cache is much smaller than that of the previous generation.
Cohere and Aleph Alpha sign merger agreement
On 16 September 2026 Cohere and German AI company Aleph Alpha signed a merger agreement, reportedly valuing the combined company at about $20 billion. The deal still needed regulatory approval.
Character.ai post-trains its own CAI-Image models on Qwen-Image
Character.ai's CAI-Image family consists of image models post-trained from open-source Qwen-Image and tuned for character-driven storytelling, so the same character stays recognisable across styles, scenes and pages. The company announced the models on 16 September 2026.
Character.ai's CEO joins Disney as its CTO
On 18 September 2026 Character.ai said its CEO Karandeep Anand was joining Disney as its Chief Technology Officer, a newly created role, with members of Character's technical team expected to follow.
SpaceXAI releases Grok 4.7
On 21 September 2026 xAI released Grok 4.7, with a larger base model and a longer reinforcement-learning run. Its Fast variant was limited to Cursor and Grok Build.
Anthropic releases Claude Opus 5.5
Released on 22 September 2026, Claude Opus 5.5 is the first model of the Claude 5.5 family. Anthropic said it matched Claude Fable 5.1 on most work, at a lower price.
PixVerse releases R2 and turns its world model into playable games
On 22 September 2026 PixVerse announced R2, a scaling release of its real-time world model that adds more data, tasks and input types at the same latency, and showed it running playable, prompt-editable worlds.
Anthropic releases Claude Sonnet 5.5
Released on 28 September 2026, Claude Sonnet 5.5 is the second model of the Claude 5.5 family. It kept the price of Sonnet 5 and was the first Sonnet model to ship with cyber safeguards and anti-distillation classifiers.
Manus 2.0
On 28 September 2026 Manus released Manus 2.0, presented as a new architecture and new products: an in-house agent harness called Cascade, a purchasable Cloud Computer, event-triggered automations, and a shared workspace, Manus Studio.
ElevenLabs launches Eleven v4, its most expressive text-to-speech model
On 28 September 2026 ElevenLabs released Eleven v4, a text-to-speech model it calls its most emotive yet, alongside a low-latency Turbo variant; the company says it topped Artificial Analysis and was preferred by about 75% of listeners in blind tests.
OpenAI releases GPT-6.1 Sol at DevDay 2026
On 29 September 2026 OpenAI released GPT-6.1 Sol at its DevDay developer event, saying it nearly matched GPT-6 Astra at lower cost. A planned Astra release was withheld over safety results, and GPT-6.1 Terra followed in October.
Moonshot AI releases Kimi K3.1 Pro
In October 2026 Moonshot AI released Kimi K3.1 Pro, an open-weights upgrade for financial research, 3D modeling and coding. It was the first model in the K3.1 series.
MiniMax releases MiniMax-M3.1-Flash
MiniMax posted the weights of M3.1-Flash to Hugging Face on 17 October 2026. The model is a multimodal M3-line release with about 428 billion parameters, MiniMax Sparse Attention throughout, and DSpark speculative decoding.
Waymo announces its first Canadian service, in Toronto
Waymo said it will launch its autonomous ride service in Toronto, its first Canadian city, with the first passengers planned for 2028.
Aleph Alpha releases the Apache-2.0 model Kolibri
On 3 October 2026 Aleph Alpha released Kolibri, an open-weight model pitched at public administration, industry and aerospace, with German-language specialization and a traceable training process presented as its advantage rather than raw scale [^1].
Mistral AI unveils Mistral Large 4, a trillion-parameter open-weight model
On 6 October 2026 Mistral AI opened a public preview of Mistral Large 4, a model with about 1 trillion total parameters. On 29 October 2026 it released open weights of an improved version.
Anthropic releases Claude Haiku 5.5
Released on 7 October 2026, Claude Haiku 5.5 is the third model of the Claude 5.5 family. It is the first Haiku model with an adjustable effort setting, and its prices start at $0.10 per million input tokens.
Character.ai files for Chapter 11 to sell the company
On 12 October 2026 Character.ai began a voluntary Chapter 11 case in Delaware to carry out a court-supervised sale of the company, warning that the platform will be wound down if no actionable bid emerges.
Google launches Gemini 4 Argon
Google gave Gemini 4 Argon to trusted security partners in September 2026 and released it worldwide on 13 October, opening its newest frontier model to developers, enterprises and subscribers while naming Gemini 4 Carbon as the next model in the family.
Physical Intelligence opens the Pi API beta
On 13 October 2026, Physical Intelligence opened a beta of the Pi API, a service that lets customers upload their own robot data, fine-tune the company's models and run cloud-hosted inference on their own machines.
Meta releases the open-weight Muse Spark Open
On 15 October 2026 Meta released Muse Spark Open, an open-weight, natively multimodal model with 816 billion total parameters, optimised to run on Apple's Mac Studio. Its weights were published on Hugging Face.
Kuaishou rolls out Kling 4.0
On 15 October 2026 Kuaishou's video generation model Kling 4.0 became available to all users, adding longer shots, more dynamic motion and Omni Reference, with a faster Flash variant released alongside it.
IFM and Inception42 release the six-model Jais 3 family
On 15 October 2026, IFM and Inception42 announced Jais 3, six Arabic-centric language models from 900 million to 375 billion parameters plus a speech recogniser, evaluated on 24 audited Arabic benchmarks.
DeepSeek updates V4.1-Flash and opens its Search and Fetch APIs
On 17 October 2026 DeepSeek shipped DeepSeek-V4.1-Flash-1017, an update to its smallest model with better benchmarks, faster task completion and shorter outputs; its weights were published and its Search and Fetch APIs opened at $0.75 per 1,000 calls.
Liquid AI adds an omni-modal model to the LFM3 family
On 21 October 2026 Liquid AI released LFM3-Omni-2B, the first omni-modal model in its LFM3 family, taking text, images, video and audio as input and returning text or speech. At 2.17B parameters it targets phones and laptops.
Anthropic donates Agent Skills to the Agentic AI Foundation
On 22 October 2026 Anthropic handed the Agent Skills open standard — its name, logo, GitHub organisation and domain — to the Agentic AI Foundation, a Linux Foundation project, after more than 70 agent platforms had adopted it.
iFlytek releases device-cloud Spark X2.6
On 22 October 2026 iFlytek released Spark X2.6, a device-cloud model series, at its developer festival in Beijing. The same event also launched a speech foundation model and a digital-employee platform.
AAIF adds Agent Skills, Open Responses, OpenEnv and OpenSandbox
On 22 October 2026 the Agentic AI Foundation added four projects — Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba — bringing it to 10 hosted projects covering different layers of the agentic stack.
The Agentic AI Foundation passes 300 members in under a year
The Agentic AI Foundation reported 305 member organisations in under a year, saying agents are moving from developer experiments to institutional infrastructure.
Anthropic files publicly for an IPO
On 26 October 2026 Anthropic filed an S-1 registration statement with the US SEC for a proposed IPO, intending to list on Nasdaq under the symbol "ANT".
PFN releases a 53-language translation model built with NICT
On 26 October 2026 PFN released the weights of PLaMo 3 Translate, a 53-language translation model built on a base model from joint research with NICT, in four sizes from 0.8B to 31B and several quantised formats.
Google announces Gemini 4 Carbon
Google's frontier model announced on 28 October 2026, released first to cyber defenders and reporting leading scores on intelligence, coding and cybersecurity benchmarks.
NVIDIA releases Nemotron 3.5 Transcribe and Granary 2
On 28 October 2026 Nvidia released Nemotron 3.5 Transcribe, an open-weight, free speech recognition model covering up to 40 languages, alongside the second version of its Granary speech dataset.
AGIBOT Innovation Day 2026
AGIBOT's Innovation Day 2026 in Shanghai on 29 October 2026 introduced the Worken output metric, the AHA3 humanoid hardware architecture, the G2 Max industrial robot, Link-U OS 1.0, the GO-3 Preview and the WITA-Omni interaction model.
Anthropic says it may be close to recursive self-improvement
In a revised S-1/A, Anthropic added risk factors on recursive self-improvement, saying internal expectations suggest it may soon meet its policy threshold for dramatically accelerated AI-driven progress, while no current model has been designated as exceeding it.
Marvell agrees to buy Lumentum in a $115B all-stock deal
Marvell agreed on 2 November 2026 to buy optical chip and networking equipment maker Lumentum in an all-stock deal valued at $115 billion including debt, one of the largest technology acquisitions on record.
Six AGIBOT G2 robots run a 24-hour convenience store trial
Six AGIBOT G2 wheeled humanoid robots worked a 24-hour shift in an operating Shanghai convenience store, completing 1,730 tasks on a livestream, with four of the six passing the final assessment.
Apple requires 18+ ratings for dating and AI companion apps
On 2 November 2026 Apple told developers that apps with dating or romantic features, including AI companion apps, must carry an 18+ rating, with new content descriptors required from 16 November.
Fal doubles its valuation to $16B on $1.3B round
Fal, a startup that hosts and runs AI video, image and audio models, raised $1.3 billion at a $16 billion valuation, double its valuation from earlier in the year.
Meta acquires AI coding startup Conductor
In early November 2026 Meta acquired Conductor, a startup whose platform runs multiple AI coding agents in parallel workspaces. Conductor's cloud service was set to shut down on 20 November 2026.
Baidu releases the open-weight ERNIE 5.5 Flash Base checkpoint
On 5 November 2026 Baidu introduced ERNIE 5.5 Flash Base, an open-weight 280B-parameter mixture-of-experts pre-training checkpoint published on Hugging Face under Baidu's own license.[^1][^2]
SenseTime launches SenseNova 6.8 Flash
On 5 November 2026 SenseTime announced the official version of SenseNova 6.8 Flash, a multimodal agent model for office and design work that offers free trial access.
Court approves bidding procedures for the sale of Character.ai, with Disney as stalking horse
A US bankruptcy court approved the bidding procedures for the sale of Character.ai's assets, with a 13 October 2026 agreement with Disney Worldwide Services as the stalking horse and an auction set for 1 December 2026.
OpenAI launches ChatGPT Cyber for vulnerability hunting
On 5 November 2026 OpenAI released ChatGPT Cyber, a model for cybersecurity work, limited to the top tier of its Daybreak Red programme. OpenAI says it scanned the Linux kernel and found flaws in Firefox.
Claude adds a Personal Finance tool
On 5 November 2026 Anthropic launched Personal Finance for Claude, a read-only tool that connects bank and credit card accounts through Plaid so users can ask questions about their money in everyday language, with no ads or sponsored recommendations.
Agility Robotics to go public through a SPAC merger
Agility Robotics announced it will go public through a SPAC merger, raising $465 million as it works to commercialize Digit robots and launch the next-generation Digit 5.
OpenAI reportedly raises $11B at a $1.25T valuation
Bloomberg reported on 6 November 2026 that OpenAI had raised $11 billion so far in an unclosed round valuing it at $1.25 trillion, below the $30 billion it reportedly sought.
