Mistral has released the weights for Mistral Large 4, its trillion-parameter open model, letting companies and governments download and run the system on their own infrastructure. The French AI company, based in Paris, first showed the model in preview earlier this month and put the final version on Hugging Face on Thursday, according to TechCrunch.
The release matters because it targets cybersecurity, where organizations want direct control over their data and defenses. Mistral argues that open weights give enterprises and public institutions frontier-level AI that cannot be cut off by an outside provider, a position CEO Arthur Mensch summed up by saying AI should be "a tool for empowerment, not dominance."
Mistral Large 4 specs and benchmark scores
Mistral says Large 4 reaches state-of-the-art performance at a fraction of the compute and capital spent by closed labs such as OpenAI and Anthropic. The company trained it entirely in Europe and pitches it as "sovereign" AI that can run on a customer's own servers or through platforms such as Microsoft Foundry.
Independent testing paints a more mixed picture. The Decoder reports that on Artificial Analysis' new Intelligence Index version 5, Large 4 scores 26 points, behind Chinese open models DeepSeek V4.1 Flash and GLM-5.3, both at 32, and MiMo-V2.6-Pro at 30. Mistral disputes the framing and says its model ranks first among open models on the Artificial Analysis Coding Agent Index, though it does not include Moonshot's Kimi K3.1 Pro, whose weights were due by November 4, 2026.
- About 1 trillion parameters in total, with 52 billion active per step.
- Nicknamed "Le Chonk"; takes text and images in, outputs text and code.
