Michael JohnsonOn October 29, Mistral released weights for the updated Mistral Large 4, known as Le Chonk. Developers and organizations can now download and fine-tune the model, then deploy it on their own infrastructure instead of having to use Mistral’s service. There is an important caveat for businesses: offering the model as a Model-as-a-Service requires a separate commercial license.
What’s now available
In its announcement of the weight release, Mistral said the version had been improved since the public preview began in early October. The company says the model can be tried through Mistral Studio and its API, while weights for self-hosting can be downloaded from Hugging Face. It names vLLM and SGLang as deployment options; an NVIDIA NIM container and hosted NVIDIA endpoints are also available.
The practical change is the ability to control where the model runs and adapt it to specific tasks. A team can deploy it in its own environment, tailor it to internal code or data, and manage the integration itself. This is particularly relevant to organizations that need to keep data processing within their own infrastructure. Managed options remain available through Mistral API, Mistral Studio, OpenRouter, and Microsoft Foundry.
What the published benchmarks show
Mistral reports results on several benchmarks, but these are the company’s own figures and selected evaluations. The model scored 19.7% on the private Cognition FrontierCode 2.0 benchmark and 48% on the Artificial Analysis Coding Agent Index, with an average token cost per task that the company says is 47% lower than that of the leading open-weight model in its comparison. On Terminal-Bench 4.0, the score rose by 13 percentage points from the preview to reach 41%.
For cybersecurity specialists, Mistral cites a score of 51% on the Artificial Analysis Cyber Index, a benchmark that evaluates vulnerability detection, reproduction, and remediation. The company also reports an average refusal rate of 96.6% on malicious cyber requests from JailbreakBench, StrongREJECT, and AgentHarm. These figures reflect specific task sets; organizations choosing a model should also evaluate it on their own tasks and safety settings.
Who gets more choice from the release
Small development teams can explore and adapt the model without building their own foundation model. Enterprises can decide where processing takes place, which internal data to use for fine-tuning, and how to integrate the model into existing systems. Research and security teams can study its behavior and test it in controlled scenarios.
However, “open weights” here means access to the model parameters and the ability to use them under a license, not necessarily free operation. Local deployment requires suitable computing infrastructure and deployment work. Mistral does not include a full technical report in this announcement; the company says it will publish one later, so the reported results should be considered in the context of the tests described.
What to check before deployment
The Mistral Ecosystem Open-Weight license allows users to use, modify, fine-tune, and commercialize the model. The exception applies to companies that offer it as a Model-as-a-Service: that kind of commercial offering requires a separate Mistral license. Before deployment, check the planned product against the model card and license, then test quality, speed, and cost on tasks typical of your team.
In short, the main outcome of the release is not a promise that everyone will benefit equally, but greater control: Le Chonk’s weights can now be obtained and run independently, while cloud services remain an option. For teams that value model customization and control over their environment, this is a concrete new choice; license terms and the cost of running their own infrastructure will determine how well it suits a particular project.
