Mistral puts a 1-trillion-parameter model in public preview

Mistral said Tuesday in a blog post that a public preview of Mistral Large 4 is live on Mistral Studio, and that the weights will drop by the end of October.

The Paris lab’s unofficial nickname for the model is “le Chonk.” It is a mixture-of-experts system the company describes as natively multimodal, with 1 trillion parameters in total and 49 billion active on a given token (company docs list 52 billion active and 1.05 trillion total). Customers can hit the preview API now. The weights are not downloadable yet.

How it was trained: Mistral says it trained Large 4 from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers, and that the preview runs on that same stack. A large share of the training data covered more than 160 languages, including every official language of the European Union, the company says.

What it costs on the API: Mistral’s model docs at docs.mistral.ai list preview pricing of $0.68 per million input tokens and $2.09 per million output tokens after the discount shown on that page. The company blog post does not name a price. The license for the forthcoming weights is not stated in the post.

The cyber pitch: On one Artificial Analysis Cyber Index test — reproduce a real vulnerability in open-source software, then patch it — Mistral says Large 4 scores 82%, the highest of any model. It says Claude Opus 5.5 and GPT-6 Astra score near zero on that same test because they refuse the task. On Cybench, a set of 40 capture-the-flag exercises, Mistral says the model solves 93%. Those scores and the refusal claim are Mistral’s.

Mistral says vetted cybersecurity partners and state authorities will see a build with reduced moderation during the preview. General customers get the public API.

Coding numbers: Mistral reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, and a combined Coding Agent Index of 49.8% that it uses to claim a lead over DeepSeek V4 Pro 0813 and Qwen3.8 Max. In a blind Surge AI coding review, the preview ranked second of five models at 3.74 on a 1–5 scale, behind Claude Opus 5 at 4.22 and ahead of Kimi K3 at 3.59, Mistral says.

What’s next: Mistral says the weights are due by the end of October, with more on architecture, benchmarks and post-training to come with that release.

— Rowan Keel