Mistral releases open weights for 1T-parameter Large 4 'Le Chonk', tying GLM-5.3 on Terminal-Bench

Mistral Large 4, internally nicknamed 'Le Chonk', is now fully open-weight. It has 1 trillion total parameters and 52 billion active per token. After a public preview earlier in October, Mistral ran additional reinforcement learning and fine-tuning and now claims top open-weight scores in agentic coding and defensive cybersecurity: 41% on Terminal-Bench 4.0, matching GLM-5.3 while using 47% fewer tokens per task, plus 93% on CyberBench and 82% on vulnerability patching.

The model offers three reasoning modes so users can trade cost against performance, and it was trained in Mistral's own European data centers. Mistral is leaning hard on the sovereignty pitch: companies can run security workloads locally without sending code to US hyperscalers. A companion Responses API, compatible with OpenAI's format, lowers the switching cost for teams already built on OpenAI tooling.

The launch drew 1,230 points and 343 comments on Hacker News. Unsloth and llama.cpp shipped day-one GGUF quantizations ranging from IQ2_XXS to BF16. Large 4 arrives in a crowded open-weight week alongside Meta's 816B Muse Spark Open and ongoing Qwen and DeepSeek releases. It also positions Mistral against Gemini 4 Carbon on cyber, though Google's model remains closed and gated.

The main caveat is licensing. Developers flagged that the new license restricts serving the model as an API to third parties, which makes it less open than Apache-licensed alternatives. Self-hosting a 1T MoE is also expensive, even with only 52B parameters active per token.