Mistral AI co-founder and CEO Arthur Mensch, 2024. Image: Slush / Wikimedia Commons, CC BY 4.0, cropped

Mistral has launched a public preview of Mistral Large 4, a 1 trillion-parameter model it officially nicknames “le Chonk”, and says it can do security work that Anthropic’s Claude Opus 5.5 and OpenAI’s GPT-6 Astra refuse to touch. The French company says it will release the model’s weights, so anyone can download and run it, by the end of October.

It is Mistral’s largest model to date, and the company’s pitch is plain: this is the strongest open-weight model built outside China, trained in Europe and able to run on a customer’s own servers.

What Mistral Large 4 is

The model has 1 trillion parameters in total, of which 49 billion are active for any one answer, and it handles text and images natively. Mistral says it was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own data centres in Europe, and that the preview runs on the same hardware. More than 160 languages were in the training data, including every official EU language.

Developers can try it now through the preview API in Mistral Studio. It will be offered in several regions worldwide, including a European deployment that Mistral says it runs “end-to-end, independently of other digital service providers and under European law”. Mistral says the reinforcement learning behind it is “still in flight”, so the model should keep improving before the weights arrive.

The security work closed models won’t do

Mistral leans hardest on cybersecurity. It says the model ranks in the top five on the Artificial Analysis Cyber Index, which measures how well models find and fix flaws in real software, and leads open-weight models built outside China “by a wide margin”.

On one of the index’s tests, which asks a model to reproduce a real vulnerability in open-source code and then patch it, Mistral says its model scores 82%, the highest of any model. Claude Opus 5.5 and GPT-6 Astra, it says, “score near zero on the same test because they refuse to perform the task”. Mistral argues that proving a flaw is real is where defence often starts, and that safety filters in closed models can get in the way.

It also claims its model refuses malicious cyber requests more often than any other open model on three public jailbreak and misuse benchmarks. Until the weights go out, Mistral is red-teaming it with security firms, vetted partners and “state authorities”, who get a version “with reduced moderation and expanded cyber capabilities”.

That is a direct challenge to the cautious approach of the US labs. Last month Anthropic warned that a Chinese open model, GLM-5.3, could build working hacks on its own, the kind of capability that can’t be taken back once weights are public.

How it compares on coding and agents

Mistral says the model scores 61.7% on DeepSWE v1.1 and 49.8% on a combined coding agent index, ahead of DeepSeek V4 Pro and Qwen3.8 Max. In a blind coding test run with Surge AI, annotators ranked it second of five models, behind only Claude Opus 5 (3.74 against 4.22 out of 5) and ahead of Kimi K3 and two versions of GLM.

All of these figures come from Mistral’s own announcement, and outside evaluators haven’t yet tested the preview in full. The launch also lands a day after Nvidia-backed Reflection unveiled Beam, another US-made open model pitched as a rival to China’s, and as Chinese labs such as DeepSeek raise fresh money. Mistral says it is the first model funded by its €3 billion Series D.

Update: the weights are out, and Mistral is tightening its guardrails

Mistral released the weights of an improved Mistral Large 4 on October 29, keeping its end-of-October promise. They are on Hugging Face under the Mistral Ecosystem Open-Weight licence, a modified MIT licence: anyone can use, change, fine-tune and sell the model, but businesses that offer it as a paid model-as-a-service must get a separate commercial licence from Mistral first. The model card lists 49 billion active parameters per token (52 billion counting embedding and output layers) and a 1 million-token context window, with setup examples for a server of eight GPUs and a smaller NVFP4 version for Nvidia’s Blackwell chips. Through Mistral’s own API it costs $0.70 per million input tokens and $2.10 per million output tokens.

Mistral says three more weeks of reinforcement learning since the preview have made it the best open-weight model for coding. It reports 19.7% on Cognition’s private FrontierCode 2.0 test, just ahead of DeepSeek V4.1 Flash (19.3%), and 41% on Terminal-Bench 4.0, 13 points up on the preview. In Surge’s blind coding comparisons, it says graders preferred it to GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, with only Claude Opus 5.5 ahead. It scores 51% on the Artificial Analysis Cyber Index and now has three reasoning settings: low, medium and high. As before, all of these figures come from Mistral.

The most striking line is an admission. Watching security-related use of the preview in live traffic, Mistral says the model refused to carry out black-box attacks on live targets, but “we see a potential for uplift to conduct real attacks”. It says it is changing its guardrails and will keep updating them “in the months to come”. Mistral puts the model’s average refusal rate on malicious cyber requests at 96.6% across three jailbreak and misuse benchmarks, the highest it measured among open models, and says a specialised variant for penetration testing and red-teaming is going only to vetted partners such as Orange Cyberdefense, for use on systems they own or are authorised to test. That is a different line from Moonshot, which says it cut cyber training data out of its coming Kimi K3.1 Pro weights rather than push the open frontier.

Why it matters

Now that the weights are out, anyone can run a model that Mistral says rivals closed systems at security work, with no company in the loop to refuse a request. For defenders that is the point. It is also what worries labs that keep their strongest cyber models behind an API. For more on the open models already out there, see our guide to using DeepSeek, Kimi and GLM for free.

Sources: Mistral AI, Mistral docs, Mistral AI: Mistral Large 4 is now open weight, Mistral Large 4 model card on Hugging Face.

Latest More Labs news

More More Labs news