Mistral releases the weights of Mistral Large 4, a 1-trillion-parameter model, at about half the preview price

Mistral published the weights of an improved Mistral Large 4 ("Le Chonk"), its flagship natively multimodal model of 1 trillion parameters, which entered public preview earlier this month. Weights are on Hugging Face as Mistral-Large-4-1T-A52B and an NVFP4 version. Mistral calls it European-trained and the best open-weight model for coding. A commenter says the license is a modified MIT, not an OSI-approved one.

Changes Mistral reports since the preview: three weeks of further reinforcement learning, three reasoning modes (Low, Medium, and High as the default; Low replaces "None"), and fewer output tokens. A Hacker News commenter reads the docs as pricing of $0.70 per million input tokens and $2.10 per million output, against $1.36 and $4.18 for the preview.

Mistral's claims (its own post, not independently verified unless noted):

• Coding: 19.7% on Cognition's private FrontierCode 2.0, which it says leads; 41% on Terminal-Bench 4.0, up 13 points from the preview with 30% fewer output tokens; 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task. In Surge Code's blind comparisons it was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, behind only Claude Opus 5.5.
• Security: 51% on the Artificial Analysis Cyber Index, and an average 96.6% refusal rate on malicious cyber requests, which it says is the highest among open models. It is tightening its guardrails after seeing live traffic, and gives some design partners vetted preview access to a cybersecurity variant.
• Work: 73.3% on Finch (finance work), 84.5% on BrowseComp, and 97.1% attack resistance on the B3 agent security benchmark (93.3% in the preview).

The post carries statements from companies including Booz Allen, Cisco, Cloudflare, CrowdStrike, NVIDIA and Hugging Face.

Artificial Analysis (a benchmark site, its own measurements) says the release, at 1.05T total and 52B active parameters, narrowly makes Mistral Large 4 the #1 open-weights model developed outside China. With high reasoning it scores 51 on its Cyber Index, third among current open-weights models behind DeepSeek V4.1 Flash 1017 (57) and MiMo-V2.6-Pro (56), and it ranks second among open-weights models on its Coding Agent Index when run in Mistral Vibe.

Serving: vLLM and SGLang both announced day-0 support. SGLang says it is the first to serve the model on NVIDIA Vera Rubin; vLLM lists FP8 and NVFP4 checkpoints with a built-in DSpark drafter.