Mistral Large 4 releases its weights to compete in AI

Mistral has released the weights for Mistral Large 4, a multimodal model with 1 trillion parameters that aims to compete with the most capable systems in programming, cybersecurity and enterprise work. The company presents it as an alternative that organizations can download, adapt and run under their own control.

The model, known internally as Le Chonk, was already available in public preview. After three additional weeks of reinforcement training, Mistral says it made particular gains in programming and defensive security, while also reducing the amount of text it needs to generate to complete certain tasks.

More programming capacity with fewer tokens

On Artificial Analysis's Coding Agent index, Mistral Large 4 scored 48%, virtually tied with the highest-ranked open-weight model, according to Mistral. The difference is efficiency: it used 47% fewer tokens per task on average.

In Terminal-Bench 4.0, a command-line programming test, it reached 41%, 13 percentage points higher than in the preview. It also reduced output tokens by 30% and took the lead among open-weight models in that evaluation.

In another private test by Cognition, which assesses whether a code change would be accepted by a repository maintainer, it scored 19.7%. Mistral says this was the model's best score on FrontierCode 2.0.

The company also adds three reasoning levels to control the balance between quality, speed and cost:

• Low: simple tasks.
• Medium: everyday use.
• High: difficult problems that require greater precision. This is the default option.

In practice, this lets you use less reasoning for a simple fix and reserve the more expensive mode for an agent that needs to modify a large project or investigate a complex bug.

Defensive cybersecurity, with limits

Mistral says Mistral Large 4 scored 51% on Artificial Analysis's Cyber Index, a test that measures whether a model can find vulnerabilities, reproduce them, validate them and fix them without breaking existing software.

The model also reached an average refusal rate of 96.6% for malicious cybersecurity requests, the highest figure Mistral says it has measured among open models. The goal is for it to help review code or test defenses without enabling attacks against systems that do not belong to the user.

The company itself acknowledges that it is still monitoring this balance. During real-world use, it observed refusals of black-box attacks against live targets, but also detected possible room to improve at real-world attacks. For that reason, it says it will adjust its security barriers in the coming months.

The use cases shown by Mistral include an agent that classifies vulnerability reports, reproduces the flaws in an isolated environment and proposes patches, as well as another that builds a web application using satellite data to find and fix its own errors.

What changes for companies and developers

The main difference is not just performance. By publishing the weights, Mistral allows a company to inspect, fine-tune and run the model on its own infrastructure, without having to depend on an external API or send its data to a provider.

The model was trained from scratch on European infrastructure, something Mistral links to its vision of technological sovereignty. It can also be customized with internal data and policies through Mistral Forge, while Microsoft Foundry offers a managed option with private deployments on Azure.

It is available through Mistral Studio, the Mistral API, OpenRouter and endpoints hosted by NVIDIA. Anyone who wants full control can download the weights from Hugging Face and deploy them with tools such as vLLM or SGLang.

The license allows users to use, modify, fine-tune and commercialize the model. The important condition is that companies wanting to offer it as a model service must obtain an additional commercial license from Mistral.

The announced price is $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. Tokens are the units of text the model processes and generates.

For you, this means Mistral Large 4 could become an option for creating coding assistants, internal tools or security systems that run on your own servers. What remains to be tested is how it performs outside benchmarks, what hardware is needed to run a model with one trillion parameters and whether its efficiency gains hold up under real-world workloads.