Foundation models

Dario Amodei is the co-founder and CEO of Anthropic. He previously worked at OpenAI, where he was Vice President of Research and co-authored the GPT-2 paper and led the research for GPT-3.

Sam Altman is the co-founder and CEO of OpenAI. He ran the startup Loopt, was president of Y Combinator from 2014 to 2019, and was removed as OpenAI’s CEO in 2023 before returning days later.

Backpropagation is the procedure that lets a multilayer neural network learn by sending output errors backward to adjust its weights, using gradient descent. A 1986 Nature paper made it widely known, and it is standard in most neural networks today.

Amazon is the American technology company founded by Jeff Bezos in 1994. Through Amazon Web Services it became a major cloud provider, and it has built its own AI models, chips and assistants, including the Nova models and Trainium chips.

A London-headquartered AI research organisation formed in 2023 by merging DeepMind with Google's Brain team, known for AlphaGo, AlphaFold and the Gemini model family.

ByteDance is a Chinese internet company founded in 2012, whose AI work includes the Doubao chatbot, the TRAE coding tool, the Seed model team and the Volcano Engine cloud platform.[^1]

Meta AI is the AI research and model-building organisation of Meta Platforms. Founded in late 2013 as Facebook AI Research, it produced PyTorch and the Llama and Muse model families, and it is separate from the Meta AI assistant product.

A residual connection adds a block's input to its output, so the block learns only the change it needs to make. This makes very deep networks trainable, and it is a standard part of today's transformers.

OpenAI is a San Francisco AI research and deployment company founded in 2015 as a nonprofit. It makes the GPT model family and ChatGPT and filed confidentially for an IPO in June 2026.

NVIDIA's first data-center GPU built on its Volta architecture, announced in May 2017. It added Tensor Cores, dedicated matrix-multiply units; OpenAI later trained its GPT-3 models on V100s.

The Transformer is a neural network architecture introduced in 2017 that relies on self-attention instead of recurrence or convolution. Most of today's large language models are built on it.

A 2018 paper by Matthew E. Peters and five co-authors that represents each word with the internal states of a pre-trained deep bidirectional language model, so vectors change with context; adding them to existing models improved results on six NLP tasks.

Pre-training trains a large model on huge amounts of unlabeled text, and fine-tuning then adapts it to a specific task with a small labeled dataset. This two-stage recipe is the basis of the term foundation model.

OpenAI's 2018 paper introducing GPT, a Transformer decoder pre-trained to predict the next token and then fine-tuned for specific tasks. It improved the state of the art on 9 of 12 tasks studied and set the recipe the GPT series followed.

BERT is a language model from Google that reads text in both directions and is pre-trained by filling in masked words. Fine-tuned per task, it set new records on question answering and language-understanding benchmarks and later went into Google Search.

Google's 2018 BERT paper introduced an encoder-only Transformer that reads text in both directions and is pre-trained by hiding words. It set new results on eleven language tasks and made download-and-fine-tune the default workflow.

An NVIDIA paper from 2019 presenting an intra-layer model parallelism method, implemented with a few communication operations in native PyTorch, that trained transformer models of up to 8.3 billion parameters and improved several benchmarks.

Scaling laws are empirical trends showing that a language model's test loss falls smoothly, as a power law, when parameters, data and training compute grow. They help plan large training runs and explain the race for compute.

A 2020 paper finding that language model loss falls predictably, as a power law, with model size, data and compute. It gave labs a quantitative reason to build much larger models and shaped the scale-up that followed.

NVIDIA's A100 is a data-center GPU, the first built on the Ampere architecture. Announced on 14 May 2020, it introduced third-generation Tensor Cores, the TF32 format and Multi-Instance GPU, and became a standard building block for training large language models.

The GPT-3 paper showed that a language model with 175 billion parameters could perform many tasks from a few examples in its prompt, with no fine-tuning, and that this ability grew with model size.

GPT-3 is an autoregressive language model with 175 billion parameters, described in a May 2020 paper. It could perform new tasks from a few examples in the prompt, without any gradient updates, and it was first offered through OpenAI's API.

Anthropic is an AI safety and research company founded in January 2021 by former OpenAI employees. It develops the Claude models and products, and its Responsible Scaling Policy sets out safety levels for its models.

InstructGPT is a family of GPT-3 models trained with reinforcement learning from human feedback so that they follow user instructions. OpenAI announced them in January 2022 and made them the default models in its API.

The NVIDIA H100 is the first data-center GPU built on NVIDIA's Hopper architecture, announced in March 2022 with 80 billion transistors, HBM3 memory, FP8 Tensor Cores and a Transformer Engine aimed at large language models.

RT-1 is a transformer-based model for real-world robot control, trained on over 130,000 episodes collected with 13 robots and reported to perform more than 700 language instructions at a 97% success rate.

A family of foundation language models from Meta AI, released in 7B to 65B parameter sizes and trained on public data; its 13B model was reported to beat GPT-3 on most benchmarks.

GPT-4 is OpenAI's large multimodal model, released on 14 March 2023. It accepts image and text input, scored near the top of human test-takers on a simulated bar exam, and was succeeded by GPT-4 Turbo, GPT-4o and GPT-5.

Gemini 1.0 is the first generation of Google DeepMind's multimodal model family, announced on 6 December 2023 in Ultra, Pro and Nano sizes; Google reported its Ultra version as the first model to beat human experts on MMLU.

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]

GPT-4o is a multimodal model from OpenAI, released in May 2024, that takes text, audio, image and video as input and generates text, audio and image output within one neural network. It was removed from ChatGPT in February 2026.

ESM3 is a protein generative model announced by EvolutionaryScale in June 2024 that reasons over sequence, structure and function together, and generated a fluorescent protein estimated to sit over 500 million years of evolution away from known ones.

Llama 3.1, released on 23 July 2024, added a 405-billion-parameter model to the Llama 3 family and extended its context to 128,000 tokens. Meta called the 405B the largest and most capable openly available foundation model.

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.

π0 is a generalist robot policy from Physical Intelligence, built by adapting a 3-billion-parameter vision-language model with flow matching so one model can output fast, continuous motor commands for dexterous manipulation.

DeepSeek-V3 is an open-weight mixture-of-experts language model that DeepSeek released in December 2024. Each token uses 37 billion of its 671 billion parameters, and DeepSeek describes its chat model as comparable to leading closed-source systems.

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.

Isaac GR00T N1 is an open foundation model for humanoid robots that NVIDIA announced on 18 March 2025. It is designed to be post-trained for a specific robot or task, and it pairs a planning model with quick motion control.

Qwen3 is a family of large language models from Alibaba, released on 29 April 2025 under Apache 2.0. Its main feature is hybrid thinking, which lets one model reason step by step or answer directly.

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.

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.

GPT-6 is OpenAI's sixth-generation model family, launched from 3 September 2026 with GPT-6 Astra, followed by Sol, Luna and Instant. Astra was the first OpenAI model to reach its internal critical cybersecurity threshold.

K2 Horizon is a family of six open language models from the Institute of Foundation Models, released in September 2026 with weights, code, training data and recipes under the Apache 2.0 licence.

Jais-3-375B-A23B is an open-weight Arabic-focused mixture-of-experts model from the Institute of Foundation Models: 375 billion parameters stored, 23 billion active per token, a 512,000-token context window and Apache 2.0 weights.

A vision-language-action model (VLA) is a robot policy that takes in camera images and a language instruction and outputs actions directly, usually by adapting a pretrained vision-language model to control.

An AI agent is a system that perceives its environment and acts autonomously toward goals. The term covers everything from a thermostat to LLM-driven software that directs its own tool use.

Quantization stores a language model's weights and activations at lower numerical precision, cutting the memory and bandwidth that inference requires. Methods such as LLM.int8() and GPTQ made very large models runnable on far smaller hardware.

AGI is the goal of AI systems that match or beat humans across a broad range of tasks. The term has no agreed definition; a framework by Morris et al. grades systems by performance level and generality.

Alignment is the problem of making a model's behaviour match what its users and designers actually intend, not just what its training objective literally rewards. Training on human preference comparisons is one practical route to it.

Hallucination is false or misleading information that a model presents as fact. It is a central failure mode of text-generating systems and a main obstacle to trusting their output.

A multimodal model handles more than one kind of input or output, such as images together with text, inside a single model. Work on CLIP and Flamingo showed how such models are trained and how they can be steered toward new tasks.

American computer scientist known for Google's core infrastructure — MapReduce, Bigtable and Spanner — and for leading work on Google Brain, TensorFlow and, from 2023, the role of Google's chief scientist.

Apple is a US technology company founded in 1976 that makes the iPhone, iPad and Mac. It introduced Apple Intelligence in 2024, builds its own Apple Foundation Models, and in 2026 agreed to build its next models with Google.

Mustafa Suleyman is a British AI entrepreneur who co-founded DeepMind and Inflection AI, and in 2024 became chief executive of Microsoft AI, leading its Copilot and in-house model work.

Greg Brockman is a co-founder of OpenAI. He led early projects such as OpenAI Gym and OpenAI Five, later became president, and took over the product organisation in April 2026.

Jakub Pachocki is OpenAI's chief scientist. He joined the company in 2017, led work on GPT-4 and OpenAI Five, and helped refocus its research on scaling deep learning systems.

Tokenization splits text into the tokens a language model actually reads; most modern systems do it with byte pair encoding, a 1994 compression method adapted for translation in 2015.

Tencent is a Chinese internet company founded in 1998, known for QQ, WeChat and video games. Since 2023 its Hunyuan model line has been its main AI effort, later branded Hy.

Baidu is a Beijing-based Chinese internet company founded in 2000. It runs China's dominant search engine and develops the ERNIE model family and the ERNIE Bot chatbot.

NVIDIA Research is the research arm of the chipmaker NVIDIA, founded in 2006 with more than 200 scientists. It leads work on AI, self-driving cars, high-performance computing and graphics, and it builds the company's open Nemotron language models.

Alibaba Cloud is the cloud computing arm of Alibaba Group and the developer of the Qwen (Tongyi Qianwen) large language model family, which is released as both open weights and proprietary models.

A technique that re-centres each layer's inputs to zero and rescales them to unit variance over a mini-batch, making deep networks faster and more stable to train.

Knowledge distillation trains a small model to reproduce the outputs of a larger model or an ensemble, so that much of the predictive performance is kept at lower compute and deployment cost.

Hugging Face is an AI platform company founded in 2016. Its Hub shares models, datasets and demo apps, its libraries include Transformers, and in September 2026 Nvidia agreed to acquire it.

A data company founded in 2016 that produces labeled and evaluated training data for AI models, counting OpenAI, Meta and Microsoft among its customers; in 2025 Meta took a 49% non-voting stake for more than $14 billion.

A component of sequence models, especially transformers, that encodes each token's position as a vector, giving order-blind attention a sense of sequence; its forms range from the original transformer's sinusoids to RoPE and ALiBi.

AI21 Labs is an Israeli AI company founded in Tel Aviv in November 2017 that builds large language models, including the Jurassic and Jamba families, and the Wordtune writing assistant.

ULMFiT is a 2018 transfer-learning method by Jeremy Howard and Sebastian Ruder for fine-tuning language models across NLP tasks; it beat the state of the art on six text-classification tasks and performed well with few labels.

GPT-1 was OpenAI's first generative pre-trained transformer, published in June 2018. It showed that one language model, pre-trained on unlabelled text and then fine-tuned, could improve results on many language tasks.

Cohere is a Toronto-based enterprise AI company founded in 2019. It builds the Command family of language models, sells them to businesses, and in September 2026 signed a definitive agreement to combine with Aleph Alpha.

MBZUAI, the Mohamed bin Zayed University of Artificial Intelligence, is a graduate AI university in Abu Dhabi, founded in 2019. It hosts the Institute of Foundation Models, which builds the K2, Jais and PAN model lines.

OpenAI's 2019 paper on GPT-2, a 1.5-billion-parameter Transformer trained on web text. It showed that a model trained only to predict the next word could perform several tasks when prompted, and its staged release sparked public debate.

GPT-2 is a large language model that OpenAI announced in February 2019. Its largest version has 1.5 billion parameters. OpenAI first withheld the full model, then released it in stages through November 2019.

Model parallelism splits one model's parameters and computation across multiple accelerators, so networks too large for a single device can still be trained. NVIDIA's 2019 Megatron-LM paper showed an intra-layer version built in native PyTorch.

T5 is a family of Google Transformer models that recasts every language task as text in, text out. Its 11-billion-parameter version set state of the art on several benchmarks, and the code, models and C4 dataset were released openly.

The Hugging Face Hub is the platform where people share machine-learning models, datasets and demo apps. By September 2026 it had more than 3 million models and 18 million users.

Google's TPU v4 is a machine-learning chip and supercomputer deployed since 2020, built into 4,096-chip pods with optically reconfigurable networking; PaLM and Gemini models were trained on it.

In-context learning is a large language model's ability to perform a new task from instructions and a few examples placed in the prompt, with no weight updates. Named and studied in the GPT-3 paper, it moved adaptation from fine-tuning to prompt writing.

A state space model represents a sequence as a hidden state that evolves with the input. S4 brought the idea into deep learning in 2021, and Mamba made its parameters input-dependent, enabling linear-time modeling of long sequences.

An image-text model from OpenAI that learns to match pictures with their captions. Trained on about 400 million image-caption pairs, it can classify images by comparing them against written prompts, with no task-specific training.

CLIP is an OpenAI model trained on 400 million image-text pairs to map pictures and sentences into one shared space, so it can classify images zero-shot from a written description.

ERNIE 3.0 is a 10-billion-parameter pre-trained language model from 2021 that trains on plain text plus a large-scale knowledge graph and fuses auto-encoding and auto-regressive networks, so one model handles both language understanding and generation.

Instruction tuning is a supervised fine-tuning stage that trains a large language model to follow user instructions; the FLAN paper showed it substantially improves zero-shot performance on unseen tasks.

GPT-3.5 is the family of OpenAI language models between GPT-3 and GPT-4, released through API models such as text-davinci-002 and text-davinci-003 and underlying the first version of ChatGPT.

A DeepMind study showing that, for a fixed computing budget, language models should be trained on far more data than was then standard; its 70-billion-parameter Chinchilla beat larger contemporaries.

BLOOM is a 176-billion-parameter open multilingual language model released on 12 July 2022 by the BigScience project, with weights available under a responsible-use license.

Constitutional AI is a training method introduced by Anthropic in December 2022, in which a model's behaviour is shaped by a written constitution of principles rather than by human labels of harmful outputs.

Direct preference optimization (DPO) is a method for fine-tuning language models on human preference data with a supervised-learning-style objective, removing the separate reward model and the reinforcement-learning loop of standard RLHF.

Skild AI is a robotics company founded in 2023 that builds a general-purpose robotic foundation model, intended to act as a single brain across many robot bodies and tasks such as manipulation, locomotion and navigation.

Chinese generative-AI company founded in March 2023 that built the Vidu video model and now describes two lines of work: world generation models and world action models.

01.AI is a Beijing-based artificial intelligence company founded in 2023 by Kai-Fu Lee. It builds open and closed large language models, including the Yi series, and the Wanzhi enterprise platform.

xAI is an AI company founded by Elon Musk that builds the Grok models and chatbot. It was acquired by SpaceX in 2026 and has since been renamed SpaceXAI and then SpaceXSI.

Mistral AI is a Paris-based company founded in April 2023. It builds open-weight and commercial large language models, the Le Chat assistant and the Vibe coding agent, and it has grown quickly through several large funding rounds.

The Segment Anything Model (SAM) is a promptable image segmentation model Meta AI released in April 2023. Given a point, a box or a mask, it outlines the object the prompt refers to.

StepFun (阶跃星辰) is a Shanghai AI company founded in April 2023 by former Microsoft staff. It builds multimodal Step-series models and has open-sourced several of them.

SenseNova is SenseTime's family of foundation models, unveiled on 10 April 2023 with a chatbot, an image generator and an avatar video generator, first offered to government and corporate clients rather than the public.

Baichuan Intelligence (百川智能) is a Beijing AI company founded in April 2023 by Wang Xiaochuan. It released open-source language models from 2023 and later focused on applying AI to healthcare.

Amazon Bedrock is AWS's managed, serverless service for accessing foundation models from several companies, announced on 13 April 2023. It became generally available in September 2023 and added Amazon's own Nova models and the AgentCore agent platform later.

SparkDesk is iFlytek's cognitive large model, unveiled on 6 May 2023 as the company's answer to ChatGPT in education and enterprise uses, built on Huawei Ascend chips and followed by rapid version updates through 2024.

Sakana AI is a Tokyo research and development company founded in July 2023 by David Ha, Ren Ito and Llion Jones, known for evolutionary model merging, projects such as The AI Scientist, and Japanese-language models named Namazu.

Meta's family of open-weight language models released in July 2023 in 7B, 13B and 70B sizes, each with a dialogue-tuned Chat variant, issued under a custom licence that allowed free research and commercial use.

Qwen, known in Chinese as Tongyi Qianwen (通义千问), is Alibaba's family of large language models. Its first open-weight model, Qwen-7B, was released on 3 August 2023, and the family has grown through many later versions.

Baichuan 2 is the second-generation open model family from Baichuan, released on 6 September 2023 in 7B and 13B sizes. Commercial use required an application and was limited by user numbers.

Hunyuan is Tencent's first proprietary large foundation model, unveiled in September 2023 and offered to enterprises through Tencent Cloud. The line was later branded Hy.

Mistral AI's first model: a 7.3-billion-parameter language model released in September 2023 under the Apache 2.0 license, built to match much larger models on standard benchmarks.

ERNIE 4.0 is Baidu's foundation model unveiled in October 2023, pitched as its most capable model and, in Robin Li's words, "not inferior in any aspect to GPT-4". It anchors Baidu's ERNIE Bot assistant and Qianfan platform.

Yi-34B and Yi-6B are open bilingual base models from 01.AI, released in November 2023 and trained on 3 trillion tokens. Yi-34B ranked first among open pre-trained models at launch.

Grok-1 is xAI's first large language model, announced on 3 November 2023 as the engine of the Grok chatbot. Its 314-billion-parameter weights were released openly in March 2024.

An open-weight sparse mixture-of-experts language model from Mistral AI, released in December 2023: 46.7B total parameters but 12.9B used per token, a 32k-token context, and results above Llama 2 70B on most benchmarks.

A mid-size multimodal model in Google's Gemini 1.5 generation, announced in February 2024, known for a context window that reached one million tokens and later two million.

The European AI Office is a body inside the European Commission that supports implementation of the EU AI Act and enforces the rules for general-purpose AI models. It was established in February 2024.

NVIDIA's project for a general-purpose foundation model for humanoid robots, announced in March 2024 and followed in March 2025 by Isaac GR00T N1, an open, customizable model released with training data and evaluation scenarios.

NVIDIA's data-center GPU for the Blackwell generation, announced in March 2024 as the successor to Hopper, built from two joined dies and paired with NVLink and FP4 support for large-scale AI training and inference.

EvoLLM-JP is a Japanese model from Sakana AI built by evolutionarily searching merges of three open-source models instead of training from scratch; the 7B result has mathematics ability none of its sources had alone.

Microsoft AI is the consumer AI organisation Microsoft formed in March 2024 to advance Copilot and its other AI products and research, led by Mustafa Suleyman. It develops the MAI family of in-house models.

Jamba, released by AI21 Labs on 28 March 2024, is an open-weights model that combines Mamba state-space layers with Transformer attention. It has 52 billion total parameters and a 256k-token context window.

Llama 3 is an open model family from Meta, released on 18 April 2024 at 8 billion and 70 billion parameters. It was pretrained on over 15 trillion tokens of publicly available data.

DeepSeek-V2 is a mixture-of-experts language model released in May 2024 with 236B total parameters, of which 21B are activated per token. It introduced Multi-head Latent Attention and used DeepSeekMoE for economical sparse training.

Yi-Large is a closed large language model with about one hundred billion parameters, released by 01.AI on 13 May 2024 and offered through its API platform; 01.AI reported top rankings for it on several public benchmarks.

ByteDance's Doubao model family, launched through Volcano Engine in May 2024, offered general-purpose, speech and image models at prices far below rivals, helping set off a price war among Chinese providers.

Qwen2 is the Qwen family's June 2024 generation, with five sizes from 0.5B to 72B parameters, context lengths up to 128K tokens, and most sizes released under Apache 2.0.

OpenVLA is a 7-billion-parameter open-source vision-language-action model trained on 970,000 real robot demonstrations, released with checkpoints and code so that other groups can fine-tune and run it on their own robots.

Safe Superintelligence Inc. (SSI) is an American AI company founded in June 2024 by Ilya Sutskever, Daniel Gross and Daniel Levy to build a safe superintelligence as its only product, without releasing anything in between.

A public benefit AI company that introduced itself alongside its ESM3 model in June 2024 and whose team joined Biohub in November 2025.

iFlytek's large language model released on 27 June 2024 and trained entirely on a domestic platform co-built with Huawei; iFlytek presented it as reaching GPT-4 level while acknowledging weaker coding and multimodal abilities.

Step-2 is a trillion-parameter language model from StepFun built on a Mixture-of-Experts architecture. A preview appeared in March 2024; the official version launched at WAIC in July 2024.

Test-time compute is the computation a model spends while answering a prompt, beyond what training already fixed into its weights. A 2024 study by Snell et al. showed how to spend it adaptively, and how it can stand in for a bigger model.

Qwen2.5 is the Qwen family's September 2024 generation. Its launch included 13 open-weight models, with up to 18 trillion pretraining tokens, 128K context and most models under Apache 2.0.

Yi-Lightning is 01.AI's flagship model from October 2024, built on an enhanced Mixture-of-Experts design. It ranked sixth overall on the LMSYS Chatbot Arena at launch and first among Chinese models.

Hunyuan-Large is Tencent's open-weight mixture-of-experts model, with 389 billion total parameters and 52 billion activated, released in November 2024 under a Tencent license.

Doubao-1.5-pro, released by ByteDance on 22 January 2025, is a closed multimodal model built on a sparse mixture-of-experts design. It was offered through Volcano Engine at low API prices.

Qwen2.5-Max is Alibaba's large mixture-of-experts model, announced in January 2025. It was offered through Qwen Chat and the Alibaba Cloud API, and its weights were not released.

Helix is Figure's vision-language-action model for humanoid control, pairing a 7-billion-parameter vision-language model with a fast 80-million-parameter action policy, trained on about 500 hours of teleoperated data and run entirely on onboard GPUs.

GPT-4.5 is a large language model OpenAI released on 27 February 2025 under the code name "Orion". It was its largest chat model at the time, priced far above GPT-4o, and was retired from general availability within months.

Hunyuan Turbo S is a fast-thinking model released by Tencent on 27 February 2025, built on a hybrid Mamba-Transformer design with 560 billion total parameters. Its weights were not released.

Ling is a family of open-weight Mixture-of-Experts language models from Ant Group's InclusionAI, released with a technical report arguing that a 300B-scale MoE can be trained without premium GPUs.

A general-purpose embodied foundation model released by AgiBot on 10 March 2025, built on a Vision-Language-Latent-Action architecture. In the company's tests across five tasks it raised average success rate from 46% to 78%.

Command A is Cohere's 111-billion-parameter model, released in March 2025 with a 256k-token context window. Cohere said it runs on two GPUs and delivers 150% higher throughput than Command R+ 08-2024.

ERNIE 4.5 is Baidu's native multimodal foundation model, released on 16 March 2025 and handling text, images, audio and video. In June 2025 Baidu open-sourced ten models from the family under the Apache License 2.0.

Llama 4 Maverick is a mixture-of-experts model Meta released alongside Scout on 5 April 2025, with 400 billion total parameters and a 1-million-token context. Its ranking on LMArena was later questioned.

Llama 4 Scout is a mixture-of-experts model Meta released on 5 April 2025. It has 109 billion total parameters, 17 billion of them active for each token, and a context window of 10 million tokens.

π0.5 is a vision-language-action model from Physical Intelligence, released in April 2025, built to generalize to unfamiliar homes and trained by co-training on many kinds of robot and web data.

A research institute of Abu Dhabi's Mohamed bin Zayed University of Artificial Intelligence, launched in May 2025, working from Abu Dhabi, Silicon Valley and Paris on open frontier foundation models including K2, Jais and PAN.

Falcon-H1 is a family of open-weight language models that runs Transformer attention and state space layers in parallel, released in sizes from 0.5B to 34B parameters under the Falcon-LLM Licence.

Aurora is a foundation model for the Earth system, trained on more than a million hours of geophysical data. It outperforms operational forecasts for weather, air quality, ocean waves and tropical cyclone tracks at far lower computational cost.

EXAONE 4.0 is LG's family of large language models, released as a 32B model and a 1.2B on-device model, merging chat usability with a reasoning mode and adding tool use and Spanish support.

A 14.74-billion-parameter Korean-English reasoning model released by NAVER in July 2025, built by compressing a larger model and training it with a four-stage reasoning recipe at a fraction of the cost of comparable models.

Qwen3-Max is Alibaba's proprietary trillion-parameter language model, released in September 2025 and served through Alibaba's API. It is the Max-tier flagship that followed Qwen2.5-Max.

LongCat-Flash is a 560-billion-parameter Mixture-of-Experts language model from Meituan's LongCat team, which activates roughly 27 billion parameters per token and was released with open weights under the MIT license.

ERNIE 5.0 is Baidu's natively omni-modal foundation model, unveiled in November 2025 and described as having 2.4 trillion parameters. It is proprietary and was made available to enterprises and consumers in January 2026.

Mistral AI's flagship model of the Mistral 3 family, released in December 2025: a sparse mixture-of-experts model with 675B total parameters, open weights under Apache 2.0, and image understanding.

Tencent's flagship model released in December 2025, offered in Think and Instruct versions, built on a 406-billion-parameter mixture-of-experts architecture with a 256K-token context window.

Solar Open 100B is Upstage's open-weight mixture-of-experts language model for English and Korean, released in December 2025 with 102.6 billion total parameters, 12 billion active per token, and a custom licence.

NVIDIA previewed Isaac GR00T N2, a next-generation robot foundation model built on a world action model architecture that the company says more than doubles success on new tasks compared with leading VLA models.

An open reasoning vision-language-action model from NVIDIA for humanoid robots, announced in March 2026 as commercially usable and trained on large public robot manipulation datasets.

EXAONE 4.5 is LG AI Research's first open-weight vision language model: a 33-billion-parameter model with a 1.2-billion-parameter vision encoder, trained on documents and text, with a 256K-token context window, released in April 2026.

π0.7 is a general-purpose robot manipulation model from Physical Intelligence that matches fine-tuned specialists on a wide range of dexterous tasks while following new language commands and recombining learned skills in unfamiliar settings.

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.

MiniCPM5 is a series of small, dense, open-weight language models from the MiniCPM team, beginning with a 1B model released in May 2026 and followed by a 2B model in September 2026, all aimed at on-device and other resource-constrained deployment.

MAI-Thinking-1 is Microsoft AI's flagship reasoning model, launched on 2 June 2026 at Microsoft Build. Microsoft said it was preferred to Anthropic's Claude Sonnet 4.6 in blind human evaluations, and that developers could fine-tune its weights for the first time.

Qwen3.8 is Alibaba's 2026 flagship line. Its open 2.4-trillion-parameter model, Qwen3.8-2.4T-A95B, released in August 2026, has 95B active parameters and a custom license.

Step 5 Preview is a preview release of StepFun's flagship foundation model: a sparse mixture-of-experts system with 600 billion total parameters, a 1-million-token context window and published weights, released in September 2026.

LFM3-Omni-2B is Liquid AI's first omni-modal model in the LFM3 family: one 2.17-billion-parameter model that reads text, images, video and audio and answers in text or speech.

GO-3 Preview is an embodied AI model that AGIBOT announced in October 2026, built on a World Reasoning Action Model; the company says it topped three embodied AI benchmarks and will fully open-source it by the end of 2026.

Ant Group is the company whose InclusionAI team released the open-weight Ling (Bailing) family of mixture-of-experts language models in 2025, including Ling-Lite and Ling-Plus, trained without premium GPUs and published under the MIT licence.

Prompt engineering is the practice of designing the text input that steers a generative AI system toward a desired output. Shared vocabulary and techniques such as chain-of-thought prompting make results easier to describe, compare and reproduce.

Long context is a model's capacity to take in and work over very large inputs — documents, images, recordings — in a single request. Systems now reach millions of tokens, though finding information inside that window is a separate skill.

A context window is the full span of text a model can refer to while producing a response, measured in tokens and including the response itself. Its size, and how evenly a model uses it, shape what the model can do.

A function applied to each node's weighted sum, giving neural networks their non-linearity. Sigmoid, tanh and ReLU shaped early deep learning; GELU now dominates transformers.

Synthetic data is training material written by a model rather than collected from people or the web. In the phi-1 experiments a small code model trained partly on model-generated "textbooks" reached competitive benchmark scores, suggesting data quality itself shapes capability.

Red teaming is the practice of actively probing a model for harmful or unwanted behaviour before deployment. A 2022 method by Perez and colleagues automated it by using one language model to generate test cases for another.

A jailbreak is an adversarial prompt that gets a safety-trained model to produce behaviour its training was meant to prevent. A 2023 analysis by Wei et al. traced these failures to competing objectives and mismatched generalization.

LG AI Research is the organisation behind EXAONE, a family of large language models it documents in technical reports and releases under its own EXAONE licence.

The Technology Innovation Institute is a research centre of Abu Dhabi's Advanced Technology Research Council that works across ten technology fields and develops the Falcon family of language models, releasing model weights, technical reports and its own licence.

Jared Kaplan is a physicist turned AI researcher who co-founded Anthropic and serves as its Chief Science Officer. He was first author of the 2020 paper “Scaling Laws for Neural Language Models.”

Alec Radford is a machine learning researcher who worked at OpenAI from 2016 to 2024. He was first author of the GPT-2, CLIP and Whisper papers, and lead author of the 2018 paper that introduced the GPT approach.

Physical Intelligence is a company that develops foundation models and learning algorithms for robots, aiming to bring general-purpose AI into the physical world. It publishes generalist robot policies such as π0 and π0.5, along with open weights and code.

Koray Kavukcuoglu is a DeepMind computer scientist who co-authored the first deep Q-network paper and later, as Google DeepMind's CTO and Google's chief AI architect, fronted successive Gemini model launches.

Jason Wei is an AI researcher and first author of papers on chain-of-thought prompting and emergent abilities in large language models; he has worked at Google Brain, OpenAI and Meta Superintelligence Labs.

Stanford computer science professor who directs the Center for Research on Foundation Models and works on making language models more open and more rigorously benchmarked.

Google, part of Alphabet, works on AI at several levels at once: it designs its own TPU chips, trains large language models such as PaLM and Gemini, and brought its research labs together as Google DeepMind in 2023.

Meituan's LongCat team released LongCat-Flash, a 560-billion-parameter mixture-of-experts language model that activates roughly 27 billion parameters per token on average and ships with open weights under the MIT licence.

Sarvam AI is an Indian company building a sovereign AI stack; its Sarvam-M, a post-trained 24-billion-parameter Mistral Small released under Apache-2.0, reports large gains on Indian-language benchmarks.

Liquid AI is an MIT CSAIL spin-out that calls itself an efficiency-first foundation model company, arguing that models should run on local devices; its LFM3-Omni-2B is an omni-modal model published in October 2026.

American entrepreneur born in 1997 who co-founded Scale AI, a data-labelling and model-evaluation company, and became Meta's chief AI officer and leader of Meta Superintelligence Labs in 2025.

A regularisation technique for neural networks: during training it randomly zeroes the inputs or outputs of some neurons so that units cannot co-adapt, which reduces overfitting.

Covariant is a robotics AI company founded in 2017 by researchers largely from UC Berkeley; in 2024 Amazon hired its founders and took a non-exclusive licence to its robotics foundation models.

An independent, non-profit AI research institute in Toronto, launched in 2017 with provincial and corporate funding, with Geoffrey Hinton among its founding scientists.

Stanford's Institute for Human-Centered Artificial Intelligence, launched in 2019 and co-directed by Fei-Fei Li and John Etchemendy, studies AI's opportunities and risks and houses the Center for Research on Foundation Models.

Langboat Technology is a Chinese AI company founded in June 2021 and incubated at Sinovation Ventures; it develops the in-house Mengzi model family and sells an enterprise agent platform and applications built on it.

Jurassic-1 is the first family of large language models from AI21 Labs, announced with the AI21 Studio developer platform in August 2021. Its largest model, Jumbo, has 178 billion parameters.

Megatron-Turing NLG 530B is a 530-billion-parameter language model announced by NVIDIA and Microsoft on 11 October 2021, then the largest monolithic transformer language model, trained on NVIDIA's Selene supercomputer with Megatron-LM and DeepSpeed combined.

ERNIE 3.0 Titan is a Chinese pre-trained language model with over 260 billion parameters, described by its authors as the largest dense Chinese model of its time and reported to beat state-of-the-art systems on 68 NLP datasets.

A Chinese AI company founded in 2022 under Tsinghua associate professor Liu Zhiyuan, known for the MiniCPM on-device model series, multimodal and speech models, and the ForgeTrain training framework.

OPT-175B is a 175-billion-parameter language model from Meta AI, announced in May 2022. Meta said it was the first model of its size released with both pretrained weights and the code to train and use them.

Cohere's first flagship instruction-following generative model, released in 2023 and distributed through Cohere's API and cloud marketplaces such as Amazon SageMaker JumpStart and Oracle Cloud Infrastructure, later superseded by Command R and deprecated in 2025.

Jurassic-2 is a family of large language models released by AI21 Labs in March 2023 as the successor to Jurassic-1, offered in three sizes with instruction-tuned variants, multilingual support and task-specific APIs; it was succeeded by Jamba in 2024.

Stable LM is Stability AI's family of open language models, first released in April 2023 as 3B- and 7B-parameter alpha versions under licenses that allow commercial use of the base models and research-only use of the instruction-tuned ones.

ERNIE 3.5 is a Baidu large language model announced in June 2023 as the upgrade behind the ERNIE Bot chatbot, adding a plugin architecture and, according to Baidu, outperforming ChatGPT on several benchmarked abilities.

Baichuan-7B is an open-source, 7-billion-parameter language model for Chinese and English released in June 2023, the first model from Baichuan, trained on about 1.2 trillion tokens and evaluated on Chinese and English benchmarks.

An open-source 13-billion-parameter language model from Baichuan, released in Base and Chat versions in July 2023, trained on 1.4 trillion tokens and benchmarked against Meta's LLaMA-13B.

Skylark, or Yunque (云雀), is ByteDance's first large language model. Released in August 2023, it powered the Doubao chatbot and was relaunched in May 2024 as the Doubao model family.

SparkDesk V2.0 is the second version of iFlytek's large language model, launched on 15 August 2023 with upgraded coding and multimodal abilities, and released alongside the iFlyCode 1.0 programming assistant.

PLaMo-13B is a 13-billion-parameter, LLaMA-based language model from Preferred Networks, pre-trained on English and Japanese open data and released on Hugging Face in September 2023 under the Apache-2.0 licence.

iFlytek's Spark V3.0, released on 24 October 2023, is a large language model the company said beat GPT-3.5 in Chinese; it arrived with a domestic Huawei-built computing platform and twelve industry models.

DeepSeek LLM is DeepSeek's first general-purpose language model series, released on 29 November 2023 in 7B and 67B sizes, each with a base version and a chat version.

Baichuan 3 is a closed language model with more than one hundred billion parameters, released by Baichuan on 29 January 2024. The company reported that it matched or beat GPT-4 on several Chinese benchmarks and on Chinese medical tasks.

Spark V3.5 is a large language model that iFlytek released on 30 January 2024. It was the first Spark model trained entirely on the domestic FlyingStar One computing platform, and iFlytek compared it with GPT-4 Turbo.

Aya 101 is a 13-billion-parameter, instruction-tuned language model from Cohere For AI that follows instructions in 101 languages. It was released under Apache 2.0 in February 2024.

Cohere's 35-billion-parameter language model announced in March 2024, built for retrieval-augmented generation and tool use in production, with a 128k-token context window and weights released for research.

Step-1 is StepFun's first language model, a system of roughly one hundred billion parameters unveiled in Shanghai on 23 March 2024 alongside the multimodal Step-1V and a preview of Step-2.

Mengzi3 is an open-source large language model series from Langboat, released in 8B and 13B sizes under Apache-2.0, built on the Llama architecture and trained further on trillions of multilingual tokens.

abab 6.5 is a mixture-of-experts language model family released by MiniMax in April 2024, reported to support a 200,000-token context and to answer all 891 queries in a long-context retrieval test.

A sparse mixture-of-experts language model from Mistral AI, released in April 2024 under the Apache 2.0 licence with 141 billion total parameters, of which 39 billion are active per token.

Octo is an open-source generalist robot policy: a transformer-based diffusion model trained on 800,000 robot trajectories that can drive a range of robots out of the box and be fine-tuned to new setups.

Yi-1.5 is an open-source family of language models from 01.AI, released in 34B, 9B and 6B sizes in base and chat forms, built by continuing to pre-train the earlier Yi models on a large new corpus.

Baichuan 4 is a closed large language model released by the Chinese company Baichuan on 22 May 2024, the company's first multimodal model and the engine behind its first AI assistant, Baixiaoying.

Aya 23 is a family of open-weight multilingual instruction-tuned models from Cohere For AI, released in 8-billion and 35-billion-parameter sizes in May 2024. It covers 23 languages and was reported to outperform Aya 101 and several contemporary open models.

ERNIE 4.0 Turbo is a large language model Baidu released in June 2024 as a faster, cheaper upgrade to ERNIE 4.0, sitting between ERNIE 4.0 and ERNIE 4.5 in the company's model line.

Apple Foundation Models (AFM, 2024) are the first generation of the models behind Apple Intelligence: a roughly 3-billion-parameter on-device model and a larger server model that runs in Private Cloud Compute.

SAM 2 is Meta's 2024 successor to the Segment Anything Model. It segments objects in both images and videos in real time, and its release included the SA-V video dataset.

EXAONE 3.0 7.8B is the first openly released model in LG AI Research's EXAONE family: an instruction-tuned model strong in Korean, published with gated weights on Hugging Face.

Jamba 1.5 is AI21 Labs' family of instruction-tuned language models, Mini and Large, which combine Transformer and Mamba layers, offer a 256,000-token effective context and ship with open weights.

OLMoE is a fully open sparse mixture-of-experts language model from Ai2: 7 billion total parameters, 1 billion active per token, pretrained on 5 trillion tokens, with weights, data, code and logs released.

Aya Expanse is a family of multilingual open-weight language models released by Cohere For AI in October 2024 in 8B and 32B sizes, covering 23 languages.

Movie Gen is a set of media foundation models Meta premiered on 4 October 2024. It covers video generation, personalised video, precise video editing and audio generation, and was not released publicly.

A large language model released by iFlytek in October 2024 and upgraded in January 2025, trained and served on the domestic Feixing No.1 computing platform; iFlytek reported it leading several Chinese and English benchmarks against GPT-4 Turbo and GPT-4o.

A family of three instruction-tuned Korean–English language models released by LG AI Research in December 2024, ranging from 2.4B to 32B parameters, with 32K-token context and weights out under a research-use licence.

Llama 3.3 70B is a text-only, instruction-tuned model Meta released on 6 December 2024. Meta said it performs close to the 405B Llama 3.1 at much lower cost and can run on a typical developer workstation.

Falcon3 is TII's family of open-weight language models from 1B to 10B parameters, released in base and instruct form; the 10B instruct model appeared on Hugging Face in December 2024 with a 32K context and a custom licence.

An open-weights 3-billion-parameter language model from Infinigence AI, released on Hugging Face in December 2024 under Apache-2.0, built on a Llama-2 architecture with a 32K-token context for local deployment.

NVIDIA Cosmos is a platform of world foundation models that generate physics-based video to train robots and autonomous vehicles, released as open-weight models in January 2025.

Mistral Small 3 is a 24-billion-parameter open-weight language model from Mistral AI, released in January 2025 under Apache 2.0 and tuned for low-latency inference on a single GPU or laptop.

Llama 4 Behemoth is a very large Meta model previewed on 5 April 2025 as the teacher for the Llama 4 family. It was never publicly released.

ERNIE 4.5 Turbo is Baidu's multimodal large language model released on 25 April 2025, presented as a faster successor to ERNIE 4.5 with lower hallucination rates and sharply cut API prices.

Sarvam-M is Sarvam AI's open-weights 24-billion-parameter model, post-trained from Mistral Small for eleven Indian languages. It runs in think and non-think modes and reports large gains on Indic, maths and coding benchmarks.

Apple Foundation Models (2025) are the second generation of Apple's models. They add image understanding and a mixture-of-experts server design, and they open an on-device model to developers through the Foundation Models framework.

MiniCPM4 is a language model built for end-side devices rather than servers, pairing a trainable sparse attention mechanism with a compact data recipe and an inference stack tuned for long-context speed.

Skild Brain is Skild AI's general-purpose foundation model for robots, trained in a universe of 100,000 different machines and meant to be fitted to existing hardware rather than built for one task.

Megrez2-3x7B-A3B is a mixture-of-experts language model from Infinigence AI built for on-device deployment, in which groups of adjacent layers share one pool of experts so that it stores 7.5 billion parameters but activates only 3 billion per token.

Microsoft AI's first foundation model trained entirely in-house, announced on 28 August 2025. A mixture-of-experts text model, it opened for public testing on LMArena and was presented as one of several specialised models Microsoft intends to coordinate.

A robot model from Physical Intelligence trained on demonstrations, human corrections and its own autonomous experience, reported to make espresso, fold laundry and assemble boxes with over 90% success.

SAM 3 is Meta's segmentation model from November 2025. It finds, segments and tracks every instance of a concept named by a short text prompt or an example image, and Meta reported it doubles the accuracy of existing systems on this task.

ERNIE 5.1 is Baidu's large language model released on 9 May 2026, derived from ERNIE 5.0's pre-training at about 6% of the usual pre-training cost, and it placed at the top of Chinese models on LMArena text and search.[^1][^3]

Apple Foundation Models 3 (AFM 3), published on 8 June 2026, is a family of five models: two on-device models, including a sparse 20B multimodal model, and three cloud models, one of which is optimised for NVIDIA GPUs.

Granite 4.2 is IBM's open-weight model generation, released on 25 August 2026 in 3B, 8B and 30B dense sizes, with a reasoning-capable 8B model under the permissive Apache-2.0 licence.

PLaMo 3 Translate is a family of translation-specialised language models from Preferred Networks, released in four sizes from 0.8B to 31B parameters, covering 53 languages and shipping in several quantised formats.

ERNIE 5.5 Flash Base is Baidu's openly published base checkpoint from November 2026: a text-only sparse mixture-of-experts model with 280B total and 14B activated parameters, released for fine-tuning rather than as a finished assistant.

Wu Yonghui is an AI researcher who spent nearly two decades at Google, worked on Google Brain and became a vice president of research at Google DeepMind, before joining ByteDance's Seed large-model team in 2025.

AI researcher who was first author of the 2016 paper introducing neural architecture search and a co-author of Switch Transformers; his X profile lists him as CTO and co-founder of Thinking Machines Lab and VP Research at Google DeepMind.

Ori Goshen is an Israeli entrepreneur who co-founded AI21 Labs in November 2017 with Yoav Shoham and Amnon Shashua and serves as its chief executive.

Jiang Daxin is founder and CEO of the Chinese AI company StepFun and a former Microsoft global vice president. He argues that scaling laws set the ceiling on model intelligence and that unifying multimodal understanding and generation is the path to AGI.

Wang Haifeng is Baidu's chief technology officer, an ACL and IEEE Fellow, and the first-listed author of the ERNIE technical reports behind the company's foundation models.

Tom Brown is an AI researcher listed among the co-founders of Anthropic and the first author of the 2020 paper that introduced GPT-3, one of the most widely cited demonstrations of few-shot learning in large language models.

Kevin Scott is Microsoft's chief technology officer, appointed on 24 January 2017 after engineering leadership roles at LinkedIn, Google and AdMob. In March 2024 Microsoft said he continued as CTO and EVP of AI, responsible for all-up AI strategy.

Daniel Gross co-founded Safe Superintelligence Inc. with Ilya Sutskever and Daniel Levy and served as its chief executive until early July 2025, when he moved to Meta's newly launched Meta Superintelligence Labs.

Mark Chen is an OpenAI researcher who led the team behind DALL·E 2 and worked on Codex; in March 2025 he moved into an expanded Chief Research Officer role.

A Chinese artificial intelligence infrastructure company that builds the middle layer letting many models run on many kinds of chips, and also releases the open Megrez model family.

Zhiyuan Liu is a tenured associate professor at Tsinghua University and co-founder and chief scientist of Beijing ModelBest, a company working to run capable large models on local devices.