Mistral Large 4 Preview and OpenAI Math Work Headline AI News

This weekly AI roundup leads with Mistral AI's public preview of Mistral Large 4, a trillion-parameter multimodal model that uses 52 billion active parameters. Mistral says the system was trained in Europe on 3,800 NVIDIA Grace Blackwell GPUs and is being pitched as a sovereign open-weight option for enterprise use. The report also covers OpenAI's mathematics research update, which includes preprints describing formal proofs for previously unresolved problems.

EXPANDED:

Mistral AI opened public preview of Mistral Large 4 on Oct. 6. The model is described as a one-trillion-parameter, natively multimodal system with 52 billion active parameters. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in European datacenters, supporting a sovereign open-weight pitch. Weights are expected by month's end. Mistral claims strong enterprise results in cybersecurity, finance, and law, plus visual grounding that may exceed some closed frontier models. Hacker News reactions ranged from praise for cost and accuracy gains to doubts about frontier parity.

OpenAI's mathematics update included preprints with formal proofs for previously open problems, such as the Unique Games Conjecture, Barnette's Conjecture, and a three-machine scheduling problem from a 1979 text. The release also retracted three Hodge-conjecture manuscripts due to a sign error. Commenters noted LLM-assisted work has now touched four of seven Millennium Prize problems, with one resolved, while P vs. NP and Yang–Mills remain open.

Count words. First paragraph: Mistral(1) AI2 opened3 public4 preview5 of6 Mistral7 Large8 4 9 on10 Oct.11 6 12. The13 model14 is15 described16 as17 a18 one-trillion-parameter,19 natively20 multimodal21 system22 with23 52 24 billion25 active26 parameters27. It28 was29 trained30 from31 scratch32 on33 3,80034 NVIDIA35 Grace36 Blackwell37 GPUs38 in39 European40 datacenters,41 supporting42 a43 sovereign44 open-weight45 pitch46. Weights47 are48 exp