French AI company Mistral has released the open weights of Mistral Large 4, a trillion parameter model nicknamed “Le Chonk,” the company said in its own announcement. The model is pitched for agentic coding and cybersecurity defense work, and businesses can now download and run it themselves.
What exactly did Mistral release?
Mistral published the weights for Mistral Large 4 on its own site on October 29, 2026, in a post titled “Le Chonk: Built for Coders, Trusted by Defenders.” The release follows a public preview of the model that Mistral opened earlier in October. Large 4 has 1 trillion total parameters in a mixture of experts design, with roughly 52 billion active at a time, and is natively multimodal.
The model ships under what Mistral calls its Ecosystem Open Weight License, a modified MIT license that allows use, modification, fine tuning and commercial use. Companies that want to resell it as a hosted service need a separate commercial license from Mistral. Mistral is also pricing API access at 0.70 dollars per million input tokens and 2.10 dollars per million output tokens.
Why is Mistral framing this as a cybersecurity model?
Mistral is positioning Large 4 for defensive security work such as finding, reproducing and patching software vulnerabilities, rather than general purpose chat. The company said its guardrails are built to tell defensive requests apart from malicious ones, and that the model refused black box attacks it was tested against on live targets during development.
Mistral also acknowledged a tradeoff that comes with open weights: the same capability that helps defenders could, in theory, be adapted by attackers, since open weights “can’t be cut off” once released. A separate, more specialized cyber variant of the model is restricted to vetted design partners only, rather than published openly.
How does Mistral say Large 4 performs?
Mistral’s own benchmark figures, which have not been independently verified, claim Large 4 ties with the Chinese open-weight model GLM-5.3 on the Artificial Analysis Coding Agent Index while using 47 percent fewer tokens per task. On the Artificial Analysis Cyber Index, Mistral reported a score of 51 percent, and said the model refuses malicious cybersecurity requests at a 96.6 percent average rate, the highest Mistral has measured among open models it tracks.
| Benchmark | Mistral Large 4 score (self-reported) |
|---|---|
| Artificial Analysis Coding Agent Index | 48%, tied with GLM-5.3, 47% fewer tokens |
| Terminal-Bench 4.0 | 41%, up 13 points from preview |
| Artificial Analysis Cyber Index | 51% |
| Malicious request refusal rate | 96.6% average |
| B3 Agent Security Benchmark | 97.1% attack resistance |
Who is backing the release?
Mistral’s announcement carries supporting statements from a long list of security and cloud partners rather than from Mistral’s own executives, including Cisco, Cloudflare, CrowdStrike, Fortinet, Zscaler, Microsoft, Hugging Face and Nvidia. Nvidia’s Kari Briski said open models “let security researchers and developers inspect, adapt and build on AI,” a framing Mistral has used to argue against depending solely on closed model providers such as OpenAI and Anthropic.
For businesses and government agencies evaluating AI tools for security work, Large 4’s open weights mean the model can run entirely on-premises, without sending sensitive code or infrastructure data to an outside API. That is likely to matter most for regulated sectors such as banking and defense, several of which are named as early partners in Mistral’s own announcement.
Mistral frames the broader push toward open weight models as “a tool for empowerment, not dominance,” arguing that businesses and governments should not have to depend entirely on closed model providers for critical infrastructure. The company has built its identity around training large models on European infrastructure, positioning Large 4 as a European alternative in a field dominated by US and Chinese labs.
Frequently asked questions
Can developers download Mistral Large 4 for free?
Yes. The weights are released under Mistral’s Ecosystem Open Weight License, which allows free use, modification and fine tuning, including commercial use. Only companies that want to resell it as a hosted service need a separate commercial license.
Is Mistral Large 4 bigger than other open models?
At 1 trillion total parameters, it is among the largest open weight models released to date. Mistral says only about 52 billion parameters are active for any given task, since it uses a mixture of experts design rather than activating the full model at once.
Is this model safe from misuse?
Mistral says its guardrails are designed to separate defensive cybersecurity requests from malicious ones, with a self-reported 96.6 percent average refusal rate on malicious requests. A more specialized cyber-focused variant is restricted to vetted partners rather than released openly.
Leave a Reply