Total 1,098
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
