Mistral launches Large 4 preview, a 1 trillion-parameter open-weight model

Mistral has launched a public preview of Mistral Large 4, which it calls ML4 and, unofficially, "le Chonk". The preview API is available now on Mistral Studio, and the company says model weights will be released by the end of the month.

ML4 is a natively multimodal mixture-of-experts model with about 1 trillion total parameters and 52 billion active ones. Mistral's documentation lists 1.05 trillion total parameters and a 1.6 billion-parameter vision encoder. Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, with a large share of multilingual data covering more than 160 languages.

Mistral makes several benchmark claims. It says ML4 is competitive with the strongest open models globally and clearly ahead of any open-weight model from the US or Europe. It reports 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4 and 59.9% on AutomationBench. In a blind Surge AI coding evaluation, it says ML4 ranked second of five models at 3.74, behind only Claude Opus 5 at 4.22.

The most distinctive claim concerns cybersecurity. Mistral says ML4 ranks in the top five models on the Artificial Analysis Cyber Index and scores 82% on a test of reproducing and then patching a real vulnerability, the highest of any model. It says Claude Opus 5.5 and GPT-6 Astra score near zero on that test because they refuse the task. These figures come from Mistral's own announcement.

That framing is also a product argument. Mistral says provider-level refusals can block legitimate vulnerability research and incident response, and that open weights let organisations run security work under their own policies, on private cloud or on-premise. Whether defenders prefer that to the safeguards of closed models is a trade-off the launch leaves open, since the same lack of refusals applies to anyone running the model.

The release also continues Mistral's push for infrastructure it controls. It says a European deployment is operated end-to-end, independently of other digital service providers and under European law, which targets customers with sovereignty requirements.

The weights are due by the end of the month, and independent testing of them will show whether the reported benchmarks hold up. Mistral says it will also publish more on architecture, benchmarks and post-training methods.

Until release, Mistral is red-teaming the model with cybersecurity leaders, vetted partners and state authorities, who get a version with reduced moderation and expanded cyber capabilities. How the public version is configured will matter.