Mistral Large 4 is a 1-trillion-parameter open-weight model

Mistral released Mistral Large 4, which it nicknames "le Chonk," a mixture-of-experts model with 1 trillion total parameters but only 49 billion active per token. That design keeps inference cost closer to a mid-sized model while the full parameter count stays behind it. The model is natively multimodal, handles more than 160 languages including every official EU language, and combines instruct and reasoning behavior in one model. The API preview is live now, with open weights promised for the end of October.

The numbers Mistral leads with are in security. On the AA Cyber Index the model ranks in the global top five and is the strongest open-weight model outside China, solving 93% of Cybench challenges and reaching 82% on a vulnerability reproduction and patching test, the highest figure Mistral reports. Coding and agent scores are solid rather than dominant: 61.7% on DeepSWE v1.1 and 59.9% on AutomationBench. On one visual grounding benchmark, Dense 200, it edges GPT-6 Astra at 42% to 41%.

One detail worth noting is where it was trained: 3,800 NVIDIA Grace Blackwell GPUs on Mistral's own European infrastructure, with a deployment option that runs end to end under European law. Pricing is $1.36 per million input tokens and $4.18 per million output.

Why it matters

If you need an open-weight model you can host yourself, Large 4's security scores and end-of-month weight release make it worth testing against Chinese open models and closed APIs, especially for teams in Europe who want data to stay under EU jurisdiction.