By Roman Aglaganov
Mistral previewed Large 4 earlier in October. On 29 October it released the weights. The French company published an improved version of its flagship, a natively multimodal model with about a trillion parameters nicknamed “Le Chonk”, on Hugging Face, so businesses and governments can run it on their own machines.
What changed since the preview
Mistral says three more weeks of reinforcement learning brought large gains in coding and defensive cybersecurity, and made the model more efficient. It now offers three reasoning settings: low for simple tasks, medium for everyday work and high, the default, for the hardest problems. Through Mistral’s own API it costs $0.70 per million input tokens and $2.10 per million output tokens.
Mistral’s claims
All the figures here are Mistral’s. It says the model scores 48% on Artificial Analysis’s Coding Agent Index, effectively tied with the leading open-weight model, the Chinese model GLM-5.3, while using 47% fewer tokens per task. On Terminal-Bench 4.0 it reports 41%, up 13 points on the preview, and on Cognition’s private FrontierCode 2.0 test 19.7%, the best of any open model. In blind comparisons run by Surge, it says, graders preferred its code to other open models, ranking it behind only Anthropic’s Claude Opus 5.5 among the models tested.
On security, Mistral reports 51% on Artificial Analysis’s Cyber Index, which tests finding, reproducing and patching vulnerabilities, and says the model refused 96.6% of malicious cyber requests in standard jailbreak tests. Security firms including Cisco and CrowdStrike, and the EU’s cybersecurity agency ENISA, are among those it says have been testing it.
Why open weights matter here
Pierre Stock, Mistral’s vice-president of science, told TechCrunch that an open model gives defenders “access that can’t be cut off by definition”. For organisations that cannot send code or logs to a US cloud provider, that is the selling point.
The licence is open, with one catch. It is a modified MIT licence that lets anyone use, change and redistribute the model, but a company that runs it as a service for others, such as through an API, must first sign a separate commercial licence with Mistral. Internal use, and use through Mistral’s own products, are exempt.
