Mistral Previews a Trillion-Parameter Model and Promises Open Weights by Month's End

Mistral AI has given itself until the end of October to deliver on the most consequential promise in its latest launch: releasing the full weights of Mistral Large 4, a model the Paris company says has about 1 trillion parameters, so that customers can run it on their own hardware.

The company opened a public preview of the model, which it calls ML4, on Oct. 6 through its Mistral Studio API. "Weights drop end of this month," the company wrote in its announcement. Until then, it said, it is red-teaming the model with cybersecurity firms, vetted partners and state authorities, who will receive a version with "reduced moderation and expanded cyber capabilities."

ML4 is a natively multimodal mixture-of-experts model with 52 billion active parameters, meaning only a fraction of its total parameters are used for any given token. Mistral called it "our largest and most capable model to date" and said it continues to improve as reinforcement-learning training, still running, progresses.

The infrastructure story may matter as much as the model. Mistral said ML4 was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers, and that the preview is served from the same facilities. It described a European deployment that it operates "end-to-end, independently of other digital service providers and under European law."

That pitch is aimed at governments and regulated industries that want AI capability without depending on U.S. cloud providers. Mistral said ML4 is the first milestone funded by its 3 billion euro Series D and that it is "significantly scaling up" compute in its own European data centers, with more capacity coming online in the coming months.

The company put cybersecurity at the center of the launch. It reported that ML4 solves 93% of the 40 challenges in Cybench, a set of exercises drawn from security competitions, and scores 82% on a test in the Artificial Analysis Cyber Index that asks a model to reproduce a real software vulnerability and then patch it. Mistral argued that open weights and self-hosting matter in security work, where a provider's refusal policy can block legitimate vulnerability research.

On software engineering, Mistral reported scores of 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4. It said ML4 resisted 93.3% of attacks on Lakera's B3 AI Security Benchmark, a measure of robustness to prompt injection, and that its refusal rate on malicious cyber prompts is higher than in earlier Mistral models. All of these figures are Mistral's own and have not been independently reproduced.

Mistral's documentation lists the model at $1.36 per million input tokens and $4.18 per million output tokens. Mistral did not say in the announcement how pricing would change once weights are public.

Several things remain open. Mistral has not published the architecture details, full benchmark set or post-training method, which it said will accompany the weights. It has not named the license under which the weights will be released, which will determine how freely companies can use and modify them. And the end-of-October date is a self-imposed deadline.

When the weights arrive, outside researchers will be able to test Mistral's claims directly. Until then, ML4 is a preview whose most important feature, open availability, is still a promise.