Mistral releases open weights for 1.05T-parameter Large 4 'Le Chonk', pitching defensive cyber

Mistral has released its largest model as open weights. Mistral Large 4 has 1.05 trillion total parameters in a mixture-of-experts design, activating roughly 52B per token, with a 1-million-token context window. It follows a public preview on October 6 and is available on Hugging Face and enterprise clouds. Mistral says it was trained in its own data centers with scaled reinforcement learning.

The pitch is cybersecurity and sovereignty. Mistral reports strong performance on cyber benchmarks while 'achieving an average refusal rate of 96.6% on malicious cyber requests, the highest we measure among open models'. Because weights are downloadable, security teams and governments can run it on-premises without sending sensitive data to US cloud providers. Mistral researcher Sagar Vaze argued that some of the most serious recent AI cyberattacks came from closed models, not open ones.

The release lands in a crowded week. Google's Gemini 4 Carbon leads closed-model cyber evals with 88.2% on SEC-bench Pro, and OpenAI's Daybreak Red reached GovCloud. Large 4 is the strongest open-weight challenger, and the Hacker News thread, 1,706 points and 438 comments, celebrated it as a major open release.

The debate is whether a capable cyber model should be open at all. Refusal rates can be fine-tuned away once weights are public. Hardware demands are also steep: a 1T MoE model needs a multi-GPU node even with 52B active. Watch independent cyber evals and quantised community releases.