Mistral releases open weights for Mistral Large 4, a 1-trillion-parameter model aimed at cyber defense

What Mistral released

Mistral AI published the weights of Mistral Large 4 on 29 October 2026. The company had opened an API-only public preview on 6 October and promised the weights by the end of the month. Mistral says the released version is improved by three more weeks of reinforcement learning, with gains in coding, defensive cybersecurity and reasoning efficiency.

According to the model card, the model is a mixture of experts with about 1 trillion total parameters, 128 routed experts of which 4 are active per token, and one shared expert. Active parameters are 49 billion per token, or 52 billion counting embeddings and output layers. It accepts text and images, outputs text and has a 1M-token context window. FP8 and NVFP4 checkpoints are on Hugging Face, and the model now has three reasoning modes: low, medium and high, which is the default. Mistral says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers.

The cybersecurity pitch

Mistral reports a 51% score on the Artificial Analysis Cyber Index, which tests whether a model can find, reproduce and patch vulnerabilities without breaking the software. It argues that defenders need open weights because provider-level refusals can block legitimate vulnerability research, and losing API access during an incident is itself a risk. In the preview announcement Mistral said several closed models score near zero on one reproduce-and-patch test because they refuse the task.

The company also says the model refused 96.6% of malicious cyber prompts drawn from JailbreakBench, StrongREJECT and AgentHarm on average, the highest rate it measured among open models. It adds that live traffic since the preview showed a potential for uplift in real attacks and that it is tightening guardrails. A separate variant with fewer restrictions, for penetration testing and red-teaming, is offered only to vetted design partners such as Orange Cyberdefense under contract. Booz Allen, Cisco, Cloudflare, CrowdStrike, ENISA and Fortinet are among the organizations quoted in the release.

Coding, price and license

On coding, Mistral reports 41% on Terminal-Bench 4.0, up 13 points from the preview, and 48% on the Artificial Analysis Coding Agent Index, which it describes as tied with GLM-5.3 while using 47% fewer tokens per task. These are vendor figures. The API price is $0.70 per 1M input tokens, $0.07 for cached input and $2.10 for output. The model is also available in Mistral Studio, on OpenRouter, through NVIDIA-hosted endpoints and in Microsoft Foundry, which Microsoft says makes it the first cloud provider to offer the model.

The license is the Mistral Ecosystem Open-Weight license. Mistral says use, modification, fine-tuning and commercial use are free, with one condition: a business that runs a Model-as-a-Service offering needs a separate commercial license. That is narrower than the plain MIT terms on some other open releases this month, such as DeepSeek-V4.1-Flash-1017.

What to watch next

Mistral says a full technical report is coming soon but gives no date. Independent results on the Artificial Analysis leaderboards will show how the vendor scores hold up, and the guardrail changes Mistral announced will matter for anyone fine-tuning the weights. The model joins a crowded month of large open releases tracked in the model release tracker and on the Open Source AI hub.

Questions people ask

Can I use Mistral Large 4 commercially?

Yes, under the Mistral Ecosystem Open-Weight license, a modified MIT license. A business that offers the model as a hosted service needs a separate commercial license from Mistral.

How much does Mistral Large 4 cost through the API?

Mistral lists $0.70 per 1M input tokens, $0.07 per 1M cached input tokens and $2.10 per 1M output tokens.

Will Mistral Large 4 help with cyberattacks?

Mistral says its guardrails separate defensive from offensive work and that the model refused 96.6% of malicious cyber prompts on average. It also says it saw potential uplift for real attacks in live traffic and is updating those guardrails.

Sources

1. Mistral Large 4 (Le Chonk): Built for Coders, Trusted by Defenders (Mistral AI, 29 Oct 2026)
2. Mistral Large 4 model card (Hugging Face, 29 Oct 2026)
3. Introducing Mistral Large 4 (Mistral AI, 6 Oct 2026)
4. Welcome Le Chonk: Mistral Large 4 on Hugging Face (Hugging Face, 29 Oct 2026)