Frontier
Frontier
76 stories · page 1 of 9Mistral launches a preview of Large 4, a 1 trillion parameter model, and promises the weights by the end of the month
The preview is available through the API in Mistral Studio from 6 October. Mistral says the model was trained on 3,800 GPUs in its own data centers in Europe and that it will release the weights by the end of the month.
Read →OpenAI put new math results from an internal model on GitHub, and many of the proofs are formalized in Lean
On 6 October OpenAI published a repository of new mathematical results produced by an internal frontier model. Many of the proofs come with Lean formalizations that a computer can check step by step. Which model it is, OpenAI does not say; the results are listed in the repository.
Read →Anthropic puts its cyber capabilities on three access tiers, and vets applicants for the third together with the US government
Anthropic merged Project Glasswing and the old Cyber Verification Program into one program with three tiers for vetted security professionals: Defense, Red Team and Specialized. For the first Anthropic aims to answer within a few days; for the third, with the fewest blocks, it currently reviews every organization together with the US government.
Read →Google released EmbeddingGemma 2: an open 740-million-parameter model for searching text, code, images, video and audio on the device itself
One shared “shelf” for five kinds of content, which can run entirely on the device. By Google’s figures the full multimodal version, compressed, needs about 567 MB of memory on a Pixel 11 Pro. The licence is Apache 2.0, and the weights are on Hugging Face and Kaggle.
Read →OpenAI will test ads in ChatGPT during image generation and says the answers stay untouched
On 5 October OpenAI announced a new visual ad format. It will first be tested while ChatGPT generates an image, later this month in the US, with an initial group of advertisers. The ads will be labeled and separate from the image. The effect numbers it shows are partners' data.
Read →Anthropic pledges $100 million to train 10,000 engineers by the end of 2027, and the first cohorts include Accenture, Deloitte and McKinsey
On 2 October Anthropic launched Claude Frontier Academy: a commitment of one hundred million dollars and a target of 10,000 Frontier Deployed Engineers. The first programme follows the medical model: a simulated deployment, a graded task, 12 weeks of a real project at the firm and a second assessment. You get in by nomination only.
Read →Ai2 open-sources AstaBrief 8B, the model behind Asta's Fast mode: a cited report about 3.5 times faster than the Claude mode, per Ai2
Ai2 has published the model and its training data. From a research question and excerpts from the literature, AstaBrief writes a cited report, and in Asta it runs as Fast mode next to the Claude-powered Thinking mode. Ai2 itself says the training and evaluation were done mostly in 2025 and that the full evaluation has not been rerun against today's leading models.
Read →Ai2 releases Olmo-core 3: open code for training MoE models up to a trillion parameters
The Allen Institute has shown a rebuilt training system for mixture-of-experts models. Per Ai2, on 8 NVIDIA B300 chips a 47 billion parameter model trains 2.7 times faster than with their old code, and the largest test run passes 1.2 trillion parameters on 512 GPUs. They ship a report, code and a demo. No models.
Read →Google launches Guided Vision in Gemini Live: the camera says out loud what is in front of a blind user
You point the phone, Gemini tells you what it sees and how to move the frame. The feature is for blind and low-vision people and works on Android 9 and up wherever Gemini Live is available. Google states plainly that it is not a medical device and does not replace the white cane.
Read →