Quick answer: Mistral released updated weights for its Large 4 model on October 29, making the roughly 1-trillion-parameter system downloadable for organizations that want to deploy it themselves. But “open weight” does not mean lightweight or unrestricted: the model requires serious accelerator capacity, and Mistral’s license requires a separate commercial agreement for some hosted AI services.

What Mistral released

Mistral Large 4—nicknamed “Le Chonk”—was already in public preview earlier this month. The October 29 update releases improved weights after three weeks of additional reinforcement learning, which Mistral says improved coding, reasoning efficiency and defensive cybersecurity. The weights are available through the Mistral-owned Hugging Face model page; the model is also served through Mistral’s API. Mistral’s announcement describes the change as a weight release, not the debut of a brand-new model.

Why the model is called “trillion-parameter”

The official model card describes Large 4 as a multimodal Mixture-of-Experts model with about one trillion parameters in total. It says 49 billion parameters are active for each token, or 52 billion when embeddings and output layers are included. Its design routes each token through a subset of the model’s experts. That helps manage the computation, but it does not turn the full downloadable checkpoint into a small laptop model. The model card lists serving options including vLLM, SGLang and llama.cpp, and points to a smaller-footprint NVFP4 checkpoint intended for NVIDIA Blackwell GPUs.

What the coding and cyber scores do—and do not—show

Mistral reports that the updated model scores 48% on Artificial Analysis’ Coding Agent Index, effectively matching the leading open-weight model while using 47% fewer tokens per task on average. It also reports a 51% score on the Artificial Analysis Cyber Index and a 96.6% refusal rate on malicious cyber requests. These are figures from Mistral’s October 29 release. TechCrunch independently confirms the weight release and reports Mistral’s comparisons, but does not present a separate re-test of the updated checkpoint. Artificial Analysis’ public write-up on Large 4 currently concerns the earlier preview, so it should not be mistaken for independent validation of the newly released weights. TechCrunch’s report and Artificial Analysis’ earlier preview analysis provide useful context.

The distinction matters: benchmark results can help compare models, but they do not guarantee better code in a particular repository or safe handling of a live security incident. Organizations still need to test the model against their own tasks and keep human review in place.

The business catch: hosted access has a separate license rule

Mistral calls the terms the “Mistral Ecosystem Open-Weight License,” a modified MIT license. It permits users to copy, modify, deploy and distribute the model, subject to its conditions. One condition matters for AI product builders: a company operating a Model-as-a-Service business that gives third parties access to the model for inference or post-training—such as through a hosted endpoint or API—must obtain a separate commercial license from Mistral. The license excludes internal use and certain products where model capabilities are embedded within specific features or harnesses. Read the license before planning a hosted service.

That makes the release especially relevant to companies seeking to keep model deployment inside their own infrastructure. It also means startups should not assume that downloading the weights automatically grants permission to run a general-purpose hosted API for customers.

What businesses should watch next

The weight release gives organizations more control over where the model runs and how it is customized. The trade-off is that they take on infrastructure, security, evaluation and maintenance work that a hosted provider would otherwise manage. The practical questions are whether the updated model’s benchmark gains hold up in independent testing, what a reliable deployment costs, and how the license applies to each intended use.

Frequently asked questions

Is Mistral Large 4 open source?

Mistral describes it as open-weight and releases the model under its modified MIT-style Mistral Ecosystem Open-Weight License. The license permits broad use, but hosted Model-as-a-Service offerings for third-party inference or post-training require a separate commercial agreement. “Open-weight” is the more precise description of this release.

Can I run Mistral Large 4 on a regular laptop?

The full model has about one trillion parameters, so it is not a typical laptop download. Mistral provides serving guidance and a lower-memory NVFP4 checkpoint aimed at NVIDIA Blackwell GPUs, but teams should plan for accelerator infrastructure or a managed API.

What changed on October 29?

Mistral released updated weights for a model that had already entered public preview. The company says three weeks of additional reinforcement learning improved coding, reasoning efficiency and defensive cybersecurity.