Mistral Large model releases

Mistral Large 4 ("Le Chonk") is Mistral AI's flagship frontier model, launched in public preview on 2026-10-06 (realizing the "fat but sparse" open-weight MoE CEO Arthur Mensch teased for July 2026 early access). A granular Mixture-of-Experts model trained from scratch on 3,800 NVIDIA Grace Blackwell (GB200) GPUs in Mistral's European datacenters. Mistral's first-party docs list 1.05T total / 52B active parameters; some press reports 49B active, and at least one analysis (shattered.io) cites an "official" 675B/41B figure — the exact count is disputed and vendor-reported. Natively multimodal (text + image understanding via a ~1.6B vision encoder), text out, spanning 160+ languages including every official EU language, with a 1M-token context. Available now via the Mistral Studio preview API (function calling, structured outputs, document QnA, batching, Agents/Conversations endpoints) at $1.36 / $4.18 per 1M input/output tokens ($0.14 cached input); open weights are promised by the end of October 2026 (license not yet announced), so self-hosting is not yet possible. Vendor-reported benchmarks: Cybench 93%, CyberGym-E2E 82%, DeepSWE v1.1 61.7%, Artificial Analysis coding-agents index 49.8%, Lakera B3 safety 93.3%. Figures vendor-reported and unverified at launch.

Mistral's largest open-weight MoE, aimed at frontier reasoning while remaining self-hostable.

Mistral's first proprietary flagship API model, introduced alongside Le Chat and stronger multilingual/coding performance.