AI Lab
Official Mistral announcements — model releases, product launches and research, each one summarized with every source covering it, by AIBriefs. RSS
Mistral Large 4 is an open-weight multimodal MoE with 52B active and 1.05T total parameters, a 1.6B vision encoder, and a 1M-token context. Priced at $1.36/M input, $0.14/M cached input, and $4.18/M output, it is in public preview.
The collaboration pairs Mistral Large 4 with NVIDIA CUDA-X and Warp for closed-loop simulation-to-design workflows, plus Omniverse integration for 3D scene and SimReady asset generation. Mistral and NVIDIA will publish open benchmarks and datasets for physics models and agentic AI in CAE and EDA.
Voxtral 3 ships three transcription models plus one TTS model and a LiveKit-based voice agent SDK, covering transcription in 50+ languages and expressive speech in 35+ languages and 50+ dialects. Voxtral Small Transcribe 3 ranked #3 on Artificial Analysis' speech-to-text leaderboard; TTS entered the top 12.
The statement of intent targets a single pretrained Large Physics Model, the first foundation model for collider physics, trained on billions of simulated proton-proton collisions at the LHC. It would handle reconstruction, particle identification, anomaly flagging and detector-design evaluation in one model.
Timestral's largest model, at 209M parameters, ranks first on GIFT-Eval, TIME and fev-bench, beating the prior state of the art with two-thirds the parameters. Four sizes span 5M to 209M, run on standard CPUs without a GPU, and output 21 quantiles per step with up to 32k context.
Mistral Large 4 ("Le Chonk") is a 1T-parameter natively multimodal model with 49B active parameters, trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters. The preview API is live on Mistral Studio; weights drop end of this month. It scores 38 on Artificial Analysis, up from Mistral Large 3's 9.