The short answer
Mistral has made its largest model downloadable: the open weights of Mistral Large 4 went live on Hugging Face on October 29, 2026. It is a natively multimodal Mixture-of-Experts model with about 1 trillion parameters, 52 billion of them active per token, and Mistral pitches it as the best open model for agentic coding and cyber defense. For teams already evaluating Mistral models, the change is that a near-frontier model can now run fully inside your own firewall.
What did Mistral actually release?
The full model weights, not just an API. Mistral Large 4, which the company nicknames “Le Chonk,” is a hybrid instruct-and-reasoning Mixture-of-Experts model trained from scratch on European infrastructure. SiliconANGLE reports it was trained on 3,800 NVIDIA Grace Blackwell GPUs and received three more weeks of reinforcement learning after the preview. The checkpoints are published on Hugging Face in FP8 and NVFP4 formats and run on common open inference servers such as vLLM and SGLang.
The model offers three reasoning modes (low, medium and high), so teams can trade cost and latency against depth per request. Besides self-hosting, it is available through Mistral’s own API and Studio, OpenRouter, NVIDIA NIM containers and private deployments in Microsoft Foundry. That spread matters for anyone planning generative AI integration: you can prototype against a hosted endpoint and move the same model in-house later without rewriting prompts or evaluations.
How strong is it for coding and security work?
Mistral says Large 4 is the strongest open-weight model for coding. On the Artificial Analysis Coding Agent Index it scores 48%, tied with China’s GLM-5.3 while using 47% fewer tokens per task, a point TechCrunch also highlights. On Terminal-Bench 4.0 it reaches 41%, a 13-point jump over the preview, and in Mistral’s human-preference coding tests only Claude Opus 5.5 was preferred over it. On the security side it scores 51% on the Artificial Analysis Cyber Index and, by Mistral’s count, refuses 96.6% of malicious cyber requests.
Two cautions apply. Most of these numbers come from Mistral or from leaderboards it cites, and an open model that still trails the best closed systems on hard coding tasks may be good enough for code review, triage and internal tooling but not for every agent workflow. Benchmark ranks also say nothing about how the model handles your languages, frameworks and repository size.
Why does an open, sovereign model matter now?
Mistral is selling control as much as capability. TechCrunch reports the company is targeting enterprises and governments that want “sovereign cyber defense” they can run on their own systems; Mistral’s VP of Science Pierre Stock described it as access that “can’t be cut off by definition.” For buyers who have watched closed vendors change prices, retire models or restrict regions with little notice, a model you can pin to an exact version and host yourself removes a whole class of supplier risk.
The license makes that practical. Under the modified MIT terms, companies can fine-tune Large 4 and ship it inside commercial products; only providers reselling the model as a hosted service need a separate agreement. That is more permissive than several rival open releases, and it lets an internal platform team standardize on one model across code assistants, security triage and document workflows.
What it means for US & EU software teams
For EU teams, self-hosting keeps source code, prompts and logs on infrastructure you choose, which simplifies GDPR data-residency questions and DORA third-party risk reviews for financial clients. It does not make the model compliance-free: open general-purpose models still fall under the EU AI Act, and the company deploying the model stays responsible for logging, oversight and use policies. Teams that need help mapping those duties can start with an EU AI Act compliance review before rollout.
For US teams, the main effect is on cost and leverage. Mistral’s API pricing undercuts most closed frontier models, and an open checkpoint gives procurement a real fallback when negotiating with closed vendors. The trade-off is operational: a trillion-parameter model needs multi-GPU nodes, quantization choices, monitoring and on-call ownership. For many workloads a hosted endpoint or a smaller model will stay cheaper, so the decision should rest on measured cost per task, not on the appeal of owning the weights.
What should teams do now?
- Benchmark on your own work. Run Large 4 next to your current coding or security model on the same tasks and compare success rate, tokens and review time.
- Start hosted, then decide. Use the API or a private cloud deployment to test quality before buying or reserving GPU capacity.
- Price self-hosting honestly. Include GPUs, storage, engineers and on-call time, not just per-token savings.
- Have legal read the license. Confirm your use is not a Model-as-a-Service case that needs a separate commercial agreement.
- Keep the stack model-agnostic. Put prompts, evaluations and routing behind one interface so switching models stays a configuration change.
- Document governance. Record model version, data flows and human oversight for EU AI Act and client audits.
None of this is legal advice, and your obligations depend on sector and jurisdiction. But the direction is clear: open models are now close enough to the frontier that self-hosting is a serious option for coding and security work, and the teams that benefit will be the ones that test it against their own workloads first.
Frequently asked questions
What is Mistral Large 4?
Mistral Large 4, nicknamed Le Chonk, is Mistral AI's largest model to date. It is a natively multimodal Mixture-of-Experts model with about 1 trillion total parameters and 52 billion active per token, trained from scratch in European data centers. Mistral released it in preview in early October 2026 and published the open weights on October 29, 2026.
Is Mistral Large 4 free to use commercially?
Largely, yes. The weights ship under the Mistral Ecosystem Open-Weight License, a modified MIT license that lets companies use, modify, fine-tune and build commercial products on the model. The main restriction is for Model-as-a-Service providers, who need a separate commercial license before selling hosted access to the model. Legal teams should still read the license against their specific deployment.
How good is Mistral Large 4 at coding?
Mistral reports 48% on the Artificial Analysis Coding Agent Index, tied with GLM-5.3 while using 47% fewer tokens, and 41% on Terminal-Bench 4.0, up from 28% for the preview. In human-preference tests it says only Claude Opus 5.5 was preferred over it. These are vendor and leaderboard figures; performance on your own repositories needs its own evaluation.
Can we run Mistral Large 4 on our own infrastructure?
Yes. The weights are on Hugging Face in FP8 and NVFP4 checkpoints and run on vLLM or SGLang, so the model can sit inside your own data center or cloud account. A trillion-parameter model needs serious GPU capacity, so many teams will start with the Mistral API at $0.70 per million input tokens and $2.10 per million output tokens, or with NVIDIA NIM and Microsoft Foundry private deployments.
What does it mean for EU compliance?
Self-hosting keeps prompts, code and logs on infrastructure you choose, which helps with GDPR data residency and with sector rules such as DORA. It does not remove obligations: open general-purpose models still fall under the EU AI Act, and the company deploying the model remains responsible for how it is used, logged and governed.
Sources
Mistral AI — Mistral Large 4 (Le Chonk) is now open weight (primary source, October 29, 2026)
TechCrunch — Mistral releases weights for its trillion-parameter AI model, targeting cybersecurity (October 29, 2026)
SiliconANGLE — Mistral releases open weights for Mistral Large 4 (October 29, 2026)
The Next Web — Mistral releases open weights for its 1 trillion-parameter Large 4 model (October 29, 2026)