Mistral Large 4 open weights are out, built for coding

A 1 trillion parameter model you can download, with a modified MIT license and a cybersecurity focus.

Mistral AI has released the open weights of Mistral Large 4, a model with about 1 trillion parameters, after a public preview earlier in October. The company says extra training improved coding and cybersecurity. Here is what Mistral confirms, what is vendor reported, and what it means if you do not run models yourself.

Key facts

Mistral AI released the open weights of an improved Mistral Large 4 this week, following a public preview earlier this month.
It is a natively multimodal mixture of experts model with about 1 trillion parameters, support for 160+ languages and three reasoning levels: low, medium and high.
The license is the Mistral Ecosystem Open-Weight license, a modified MIT license. Free use, modification and commercial use are allowed, but companies that sell the model as a service need a separate commercial license.
Weights are on Hugging Face. The model is also on Mistral Studio, the Mistral API, OpenRouter and Microsoft Foundry.

What happened

Mistral published the weights of Large 4 after a public preview earlier this month. According to the company, three more weeks of reinforcement learning improved the model for coding and cybersecurity work. The Hugging Face listing shows two repositories, a standard version and an NVFP4 version made for Nvidia hardware.

Mistral prices its own API at $0.70 per million input tokens and $2.10 per million output tokens. Self-hosting is supported with vLLM or SGLang, and Nvidia offers a NIM container.

Most performance numbers come from Mistral itself. The company reports that the model leads open weight models on FrontierCode 2.0 with 19.7 percent, reaches 41 percent on Terminal-Bench 4.0 (up from 28 percent in the preview) and scores 51 percent on the Artificial Analysis Cyber Index. It also says the model resists 97.1 percent of attacks in an agent security test and refuses malicious cyber requests 96.6 percent of the time on average. These are vendor reported figures and were not independently checked in the coverage we reviewed.

On safety, Mistral says cybersecurity experts and government authorities got early access to a version with fewer restrictions so they could probe it. Mistral also says a security focused variant is in preview with Orange Cyberdefense, with vetting and contractual safeguards, and that it is updating guardrails because the model may still give some uplift for real attacks.

Why it matters

Most strong AI models are closed: you use them through an app or an API and cannot see or run them yourself. Open weights change that. A company can download the model, run it on its own servers and adapt it to its own data, which matters for privacy rules and for keeping costs predictable.

Mistral also says the model was trained from scratch on European infrastructure.

For everyday users nothing changes today. The effect arrives indirectly, as apps and tools built by other companies start using the model.

The license deserves a careful read. It is not plain open source: businesses that offer the model to others as a paid service need an agreement with Mistral. And because the benchmark claims come from the vendor, independent tests over the coming weeks will tell more than the launch post.

How to try it

You do not need a server to benefit from a story like this. The useful skill is turning a dense technical announcement into something you can act on. The Try it prompt, Explain simply, rewrites complicated text in plain language.

1. Open the prompt and paste it into ChatGPT, Claude or Gemini. 2. Where it asks for the text, paste the license paragraph or the availability section from the announcement. 3. Where it asks for your level, write something like “I use AI tools at work but I am not an engineer”. 4. Ask for three things: a plain summary, the one term you should look up, and one question to ask your IT team or a vendor.

Check the answer against the original. If the reply states a number or license term that is not in the text you pasted, ask the assistant to point to the sentence it came from.

Try it

Explain a topic simply

Frequently asked questions

What are open weights?

They are the trained settings of an AI model, published so anyone can download and run it. It is not always the same as open source: the training data and code may stay private, and the license can add limits, as it does here.

Can I use Mistral Large 4 for free?

The weights can be downloaded and used commercially under Mistral's modified MIT license, except for companies that offer the model as a paid service. Running a 1 trillion parameter model needs serious hardware, so most people will use it through an app or API.

Is the model safe for security work?

Mistral reports a high refusal rate for malicious cyber requests but admits some uplift for real attacks may remain and says it is updating guardrails. The security focused variant is limited to vetted partners and authorized systems.

Sources