Mistral publishes Large 4's open weights under a modified MIT license

Mistral AI released the open weights of Mistral Large 4, nicknamed Le Chonk, on Hugging Face, three weeks after the model's public preview. Hugging Face's launch post is dated October 29. Mistral describes the release as an "open-weight hybrid instruct-and-reasoning MoE with multimodal input," and the main repository is mistralai/Mistral-Large-4-1T-A52B. The model has about 1 trillion total parameters; Hugging Face's post gives 49 billion active parameters, while the repository name indicates 52 billion.

License, price and hardware

Mistral says the model "is available to use under the Mistral Ecosystem Open-Weight license, a modified MIT license." It can be used, modified, fine-tuned and commercialized, but businesses running a model-as-a-service offering must get a separate commercial license from Mistral first. Alongside the weights, Mistral set API pricing, "starting today," at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens.

Self-hosting is a data-center job. Hugging Face's examples use eight GPUs, with Inference Endpoints serving it on eight Nvidia H200s and the NVFP4 checkpoint requiring Nvidia Blackwell GPUs. 2-bit GGUF quantizations run on a single Apple Silicon Ultra machine with 512 GB of unified memory. The model supports a 1-million-token context window, and Hugging Face notes that its reasoning mode cannot be turned off, though effort can be set to low, medium or high.

The numbers Mistral reports

Mistral says the released model gained 13 percentage points on Terminal-Bench 4.0 over the preview, reaching 41%. It reports 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using about 47% fewer tokens per task, and 19.7% on Cognition's private FrontierCode 2.0, which it calls the top open-weight score. These are Mistral's own figures.

Why it matters

Open weights at this scale mainly serve companies with strict data-residency or sovereignty requirements, which can now run a model of this size inside their own infrastructure. For everyone else the more immediate effect is price. A $0.70 and $2.10 rate for a model whose weights anyone can host gives buyers a reference point when negotiating with closed-model vendors. Check the license before building a resale or hosted service on it, since that is the one use that needs a separate agreement.