Mistral releases the weights of Mistral Large 4, its 1-trillion-parameter model

Mistral Large 4, nicknamed Le Chonk, is now open weight under a modified MIT license that asks only model-as-a-service businesses for a separate deal. Where you can run it, what the API costs and which claims are Mistral's own.

Mistral announced on October 29, 2026 that it is releasing the weights of Mistral Large 4, nicknamed Le Chonk, a 1-trillion-parameter multimodal model it had put into public preview earlier in October. The weights are on Hugging Face under the Mistral Ecosystem Open-Weight license, which Mistral describes as a modified MIT license. The model is also available through the Mistral API, Mistral Studio and OpenRouter, at $0.70 per million input tokens and $2.10 per million output tokens.

The facts

  • Model: an “open-weight hybrid instruct-and-reasoning MoE with multimodal input”, in Mistral’s words. Mistral’s model card lists about 1 trillion total parameters, about 52 billion active per token (49 billion without embeddings and output layers), 128 routed experts with 4 active per token plus 1 shared expert, a 1M-token context, and text and image input with text output.
  • Weights: mistralai/Mistral-Large-4-1T-A52B on Hugging Face, plus a separate NVFP4 checkpoint for Blackwell GPUs. Mistral points to vLLM and SGLang for self-hosting; the model card’s serving examples use 8 GPUs.
  • License: free to use, modify, fine-tune and commercialize, “with a single condition: businesses that operate a Model-as-a-Service offering must obtain a separate commercial license from Mistral before using the model commercially.”
  • Reasoning modes: low (it replaces the former “None” mode), medium and high, which is the default.
  • API price: $0.70 input, $0.07 cached input and $2.10 output per million tokens. The documentation uses the model name mistral-large-4.
  • Other places to run it: NVIDIA-hosted endpoints, an NVIDIA NIM container and Microsoft Foundry.

Mistral says the model scores 48% on the Artificial Analysis Coding Agent Index, “effectively tied” with GLM-5.3 while using 47% fewer tokens per task, 41% on Terminal-Bench 4.0 (13 points above the preview) and 19.7% on Cognition’s private FrontierCode 2.0 benchmark. It also says blind graders at Surge preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, with only Claude Opus 5.5 ranked higher. These are Mistral’s figures; we have not checked them. On security, Mistral reports a 96.6% refusal rate on malicious cyber requests and also writes that “we see a potential for uplift to conduct real attacks”; it says it is modifying its guardrails accordingly.

What it means for builders

Analysis: for most builders the price and the license matter more than the download. At these rates it is a cheap model to try as a coding agent through the API or OpenRouter, and the license lets you build a product on it unless you sell access to the model itself as a hosted service. Mistral’s own serving examples use 8 GPUs, so for a small team the weights are mainly a copy you can keep running if the hosted offer changes, and a base for fine-tuning. Test it on your own tasks before you switch: the token savings lower agent costs only if they hold on your workload.

Not stated in the announcement

Hardware requirements for self-hosting, the maximum output length and a date for the technical report, which Mistral says is “coming soon”. On November 1, Mistral’s model documentation page still showed the label “Public Preview” and the date October 6, 2026.