Mistral open-sources improved Large 4 “Le Chonk” weights: open coding lead, still trails some China models overall
After preview feedback and further RL, Mistral released improved Mistral Large 4 (Le Chonk) weights. It calls the ~1T natively multimodal flagship an open coding leader on FrontierCode 2.0 and Terminal-Bench, tying GLM-5.3 on the AA Coding Agent Index with far fewer tokens. The Decoder says it still trails some Chinese open models on broader rankings.
Mistral AI announced open weights for an improved Mistral Large 4 (Le Chonk). The company says it released the ~1-trillion-parameter, natively multimodal flagship in public preview earlier this month, then used community feedback and continued reinforcement learning to lift coding and defensive cybersecurity while improving reasoning efficiency and lowering inference cost. It stresses European infrastructure training with clear lineage for sovereign, self-hosted, customizable AI.
On evals, Mistral says Large 4 leads open-weight models at 19.7% on Cognition’s FrontierCode 2.0 (end-to-end code quality and whether a maintainer would merge the change); scores 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using about 47% fewer tokens per task on average; and rose 13 points on Terminal-Bench 4.0 to 41% vs preview, with about 30% fewer output tokens. It now exposes Low / Medium / High (default) reasoning modes. Hugging Face’s blog adds: MoE with ~1T total / ~49B active parameters, 1M context, official FP8 and NVFP4 variants, a Pixtral vision encoder, and the Mistral Ecosystem Open-Weight License (MIT-derived with Model-as-a-Service limits). The Decoder instead writes ~52B active per step; both figures are publisher claims and are not reconciled here.
On cyber, Mistral reports 51% on the Artificial Analysis Cyber Index (find, reproduce, and patch vulnerabilities without breaking existing functionality) and says the model handles legitimate defensive work without unnecessary refusals, with about 96.6% average refusal on malicious cyber requests. Hugging Face’s blog separately claims 82% on an AA Cyber Index–style reproduce-then-patch test—numbers conflict with Mistral’s 51%, so both are attributed as source claims. The Decoder notes Mistral’s own chart places DeepSeek V4.1 Flash and MiMo-V2.6-Pro higher, putting Large 4 roughly third among open models, with only a one-point lift from preview (~50→51). On hardware, Decoder cites SGLang docs: standard on eight Nvidia B300 GPUs, a more compressed variant on four GB300s. Decoder’s overall take: strong open coding, but still trailing some Chinese open models on broader scores.
Sources
- Le Chonk: Built for Coders, Trusted by Defenders. — Mistral AI
- Welcome Le Chonk: Mistral Large 4 on Hugging Face — Hugging Face
- Paris-based Mistral releases "Le Chonk," a strong coding model that still trails China overall — The Decoder