Mistral releases its largest model and pitches cyber defence that nobody can switch off

France's Mistral on Thursday released the weights of Mistral Large 4, a trillion-parameter model nicknamed Le Chonk internally. Companies and governments can now download it via Hugging Face and run it on their own infrastructure, TechCrunch reports. An early version had been available through an API since the start of October. According to Heise, Mistral then tested the model with customers and cyber partners for three weeks and trained it further with reinforcement learning; the new version reportedly needs about half as many tokens for the same result.

Mistral claims the model is the best open model in the world for agentic coding and cybersecurity, and that it performs at least as well as Chinese rivals such as GLM-5.3 while using 47 percent fewer tokens per task. Heise adds caveats: on the Terminal-Bench 4.0 coding test the model scores 41 percent, roughly level with GLM-5.3 but well behind Kimi K3.1 Pro and closed frontier models. Some of the results cited by Mistral could not yet be independently verified at the time of publication.

Science chief Pierre Stock stressed to Wired that open weights cannot be switched off remotely, which he considers important for anyone arming themselves against cyberattacks. He also argued there is no guarantee that Chinese models will not behave maliciously. A researcher at security firm Sophos pointed out in Wired that freely available models can just as easily be used by attackers. Mistral attributes the ten months between its last two frontier models to building its own data centres in Europe, Heise writes.