Dashboard

Mistral Large 4 Open Weights: What You Can Ship

Mistral released the Large 4 weights on 29 October. The licence permits commercial use but carves out managed-service resale, and the benchmark table has one private row worth noting.

Cecilia Iona
Cecilia Iona
Senior Editor, AI & Product
30 October 20261 min read

Mistral published the open weights for Mistral Large 4 on 29 October 2026, moving the model from the public preview it announced earlier in the month to a downloadable release. It is a roughly 1-trillion-parameter, natively multimodal model under the Mistral Ecosystem Open-Weight licence. You can download it, fine-tune it and use it commercially. The one clause that matters: if you want to run it as a Model-as-a-Service product, you need a separate commercial licence from Mistral.

That single carve-out decides whether this release is useful to you, so it is worth getting straight before the benchmark table.

What the Mistral Large 4 open weights licence allows

Mistral describes the licence as a modified MIT licence. Use, modification, fine-tuning and commercial use are all permitted. The restriction is narrow but specific, and it lands on one business model: reselling inference.

  • Building a product that calls the model internally: allowed.

  • Fine-tuning on your own data and shipping the result inside your app: allowed.

  • Standing up an endpoint and selling access to the model itself: needs a commercial licence.

If you have read our breakdown of how open weights differ from open source, this is the familiar pattern. Weights are available, the licence is not an OSI-approved open source licence, and the constraint sits on a commercial use case rather than on redistribution in general. The clause-by-clause view in our guide to model licences and what you can legally ship applies here without modification.

The benchmark numbers, and what they are measured against

Mistral positions Large 4 as the strongest open-weight model for coding. The figures it published:

  • FrontierCode 2.0, a private Cognition benchmark: 19.7 percent, the top score in Mistral's chart.

  • Artificial Analysis Coding Agent Index: 48 percent, which Mistral puts level with GLM-5.3 while using 47 percent fewer tokens per task.

  • Terminal-Bench 4.0: 41 percent, up 13 points from the preview release.

  • Artificial Analysis Cyber Index: 51 percent.

  • BrowseComp: 84.5 percent. FinWorkBench: 73.3 percent.

  • B3 Agent Security Benchmark: 97.1 percent attack resistance, up from 93.3 percent at preview.

Two of those deserve a second look. The token-efficiency claim is the more interesting one for anyone paying per token, because a tie on quality at roughly half the tokens is a real cost difference rather than a leaderboard place. And FrontierCode 2.0 is private, which means you cannot reproduce the 19.7 percent yourself. That is not a reason to dismiss it, but it belongs in the column of numbers you take on trust. Our method for reading a vendor's benchmark table suggests starting with the rows where a vendor does not win, and Mistral does report one: in Surge Code's blind human-preference testing, Large 4 placed behind Claude Opus 5.5.

The cybersecurity framing

Mistral gave the release the codename Le Chonk and the tagline "built for coders, trusted by defenders", and a chunk of the announcement is about security work. The claim is that the model performs well on defensive tasks, vulnerability discovery, reproduction and patching, without the reflexive refusals that make security-adjacent prompts annoying on some models, while still declining genuinely offensive requests. Mistral reports a 96.6 percent average refusal rate on malicious cyber requests and says that is the highest it measures among open models.

There is also a specialised variant in preview with vetted partners including Orange Cyberdefense, which is not part of the open-weight release.

Where to get it, and whether you can run it

The weights are on Hugging Face for self-hosting with vLLM or SGLang, and NVIDIA is hosting endpoints plus a NIM container. Mistral Studio, the Mistral API, OpenRouter and Microsoft Foundry all carry it as a hosted option.

Running it yourself is a different question from being allowed to. A trillion parameters is a serving problem, not a laptop one, and the arithmetic is unforgiving regardless of how permissive the licence is. If you are weighing self-hosting against an API bill, the break-even analysis in is self-hosting an open weight model cheaper is the calculation to run first, because it turns on sustained tokens per second rather than on monthly spend.

What to do with this

If you are shipping a product that calls a coding model, the token-efficiency figure is worth testing on your own tasks before you believe it or ignore it. If you are planning to resell inference, read the licence before you plan anything else. And if you are tracking releases generally, our notes on keeping up with AI news without losing days cover how to triage announcements like this one.

How did this land?

About the author

Cecilia Iona
Cecilia Iona

Senior Editor, AI & Product

Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.

Share

Get the next post in your inbox

One email a month. Product updates, engineering posts, and the best of Built with Swarmz.

I agree to receive emails about AI building tips and Swarmz product news. Unsubscribe any time.