JOURNAL / 2026.10.31
Mistral Large 4 releases its weights: the control gained and the demands of deployment
Mistral delivers an improved Large 4 in FP8 and NVFP4. Open weights allow teams to preserve and adapt the model, while its scale, serving software and custom license determine who can benefit.
The end-of-October promise now has files behind it. Mistral announced the release of Large 4's weights on October 29, under its nickname Le Chonk. In our article about the preview, that release was still pending. We have now verified that the official repositories contain weight files accessible without prior approval. That changes an important possibility: preserving a version and deciding how to run it.
This is also more than a download of the preview. Mistral reports three additional weeks of reinforcement learning, improving coding and software defense alongside reasoning efficiency. These are vendor claims; we have not reproduced the evaluations. For this release, the strongest change is the availability of the artifact. A ranking can guide a trial; being able to pin the model lets teams repeat that trial without depending on a service update.
The main model card describes text and image input, text output, tools and three reasoning levels. It also clarifies a discrepancy from the preview: roughly 49 billion active parameters per token, or 52 billion including embeddings and output layers, within about one trillion total parameters. The distinction matters because activating a fraction of the model reduces computation, but does not make all its other weights dispensable for storage and serving.
This is where the physical constraint appears. The vLLM recipe puts FP8 weights at about 1.06 TB and NVFP4 at roughly 610 GB. These are weight sizes, not complete memory budgets. It also says support still requires a dedicated image: it is absent from ordinary releases and nightly packages. The advertised context window is around a million tokens; this recipe reports validation up to 256K. A model's maximum should not become a guarantee for every installation.
The NVFP4 card targets Blackwell GPUs and reports similar results to FP8 in its evaluations. That is a meaningful memory reduction, although published equivalence does not establish identical behavior on every document, language or tool. The Hugging Face team also documents routes through Transformers, vLLM and llama.cpp. Multiple routes expand the options; each still has its own configuration and limits.
The license distinguishes internal use from offering the model to others
The Mistral Ecosystem Open-Weight License modifies MIT. It permits use, adaptation and distribution, but requires a commercial agreement if the licensee or its affiliates operate a model-as-a-service business and want to use it commercially. Its definition turns on a third party's control over inputs, parameters or training data. It includes exceptions for internal use and access through official Mistral products, and excludes certain products with embedded features. Those distinctions need to be read against the actual architecture being offered.
My reading is that the main gain is continuity. A team able to operate it could retain the same version throughout an audit, test an adaptation and compare results afterward. That autonomy has a cost: maintaining the inference engine, providing compute capacity and reviewing behavior. Downloadable weights transfer decisions to the operator; they do not settle them.
For a small team, I would start with the API, whose current model page lists $0.70 per million input tokens and $2.10 per million output tokens. Then I would ask which concrete need justifies hosting: version continuity, adaptation or data control. I would compare both routes on the same work and count human review, waiting time and operations alongside tokens. That would be the case for deployment; having a downloadable model alone does not pay for the infrastructure.
Primary sources consulted on October 31, 2026. We checked file listings and read the license; we have not downloaded or run the weights. The adoption proposals express our judgment.