Mistral Large 4’s weights were available when checked on October 29, 2026. Mistral’s announcement releases the model, nicknamed Le Chonk, for deployment on users’ own infrastructure. The supplied announcement extract does not establish its publication date. The release comes with a business condition: the Mistral Ecosystem Open-Weight License requires a separate commercial licence for some model-hosting businesses. Internal use is explicitly exempt from that requirement.
Mistral says the release follows a public preview earlier in the month and that further reinforcement learning improved coding and defensive cybersecurity performance. Those are Mistral’s claims. No checkpoint run or independent benchmark verification accompanies this article. For teams evaluating deployment, the immediate questions are what the licence permits and which serving configuration their hardware can support.
The condition extends to affiliates
The licence text grants permission to use, modify, distribute and further train the model, subject to its conditions. Its commercial restriction applies if the licensee or any affiliate operates a Model as a Service business. In that case, a separate agreement with Mistral is required before using the model or its derivatives for any commercial purpose, subject to the explicit exceptions below. The condition is broader than charging customers for access to this particular checkpoint.
The definition turns on how much control a customer gets. Model as a Service means giving a third party access for inference or post-training while allowing meaningful control over inputs, parameters or training data. The definition excludes products whose model capabilities are embedded solely within specific features or harnesses, and services that merely relay requests to models hosted elsewhere.
Section 3 exempts internal use that does not make the model or its derivatives available to third parties, as well as use through Mistral’s official products. For a team considering a customer-facing deployment, the first distinction is whether it is offering an embedded product feature or giving customers control over a hosted model. The affiliate provision also means the operator’s wider business matters. Teams should read these provisions together before deciding whether they need a separate agreement.
Hardware recipes have different limits
Hugging Face’s published launch-guide source lists official FP8 and NVFP4 variants, alongside community GGUF conversions. It describes text and image inputs with text output, about one trillion total parameters and 49 billion active parameters. The guide explains that the model selects a small group of experts for each token. Its deployment examples still call for substantial memory and compute.
One option is an FP8 Inference Endpoints recipe using eight NVIDIA H200 GPUs. According to the guide, it splits weights across all eight GPUs and serves up to 131K tokens of context, with up to 16 requests at a time. Those limits describe that recipe. The same guide lists a one-million-token model context window, which should not be treated as the capacity of every deployment configuration.
Hardware also affects checkpoint choice. The guide’s vLLM instructions say the NVFP4 checkpoint requires NVIDIA Blackwell GPUs. A separate llama.cpp route uses two-bit GGUF quantizations on an Apple Silicon Ultra computer with 512GB of unified memory. Hugging Face reports about 15 generated tokens per second for Q2_K on a single M3 Ultra. That is its reported result for that setup. It does not establish the speed of another quantization or workload.
Mistral is asking developers to run their own evaluations now that the weights are available. For a deployment decision, pair task-quality results with the exact checkpoint, serving software and context length used. The guide’s different hardware recipes show why teams need those details before planning capacity.