Mistral releases Mistral Large 4 weights under a modified MIT license

Mistral AI published the weights of Mistral Large 4 on October 29, 2026, about three weeks after the model opened as a public API preview. The release post calls it “our flagship 1-trillion-parameter, natively multimodal model.” The weights are on Hugging Face as mistralai/Mistral-Large-4-1T-A52B, under what Mistral calls the Mistral Ecosystem Open-Weight license.

What changed

When the preview launched on October 6, Mistral promised the weights by the end of October but did not name a license. Both are now settled. Mistral describes the license as a modified MIT license: you can use, modify, fine-tune and sell products built on the model for free. The one condition, in Mistral’s words, is that “businesses that operate a Model-as-a-Service offering must obtain a separate commercial license” from Mistral.

The model card lists these details:

Mistral Large 4 weights
Size	About 1T total; 49B active per token (52B with embeddings and output layers)
Context window	1M tokens
Formats	FP8; a separate NVFP4 checkpoint for Blackwell GPUs; GGUF
Serving	vLLM (recommended), SGLang, Transformers, llama.cpp
Reasoning modes	Low, Medium and High (default), per the release post

The model card does not state minimum memory. Its vLLM and SGLang examples run on a single node of 8 GPUs, so self-hosting the full model is a data-center job, not a laptop one. The GGUF files are smaller quantized versions for llama.cpp-compatible apps.

Where you can use it

Mistral says the model is available “today” in Mistral Studio, on the Mistral API, on OpenRouter, in Vibe Code, through NVIDIA-hosted endpoints and on Microsoft Foundry, which it calls the first cloud provider to offer it. Mistral’s model docs still label the API model (mistral-large-4-0, version v26.10) as a public preview. The release post lists API prices; see current per-token prices on our AI models page.

Mistral reports 41% on Terminal-Bench 4.0, up from 28% in the preview, and 84.5% on BrowseComp. These are Mistral’s own and partners’ numbers. The company also says it saw “a potential for uplift to conduct real attacks” in cybersecurity tests. It is changing the guardrails and offers a restricted cybersecurity variant only to vetted design partners.

What to do

If you were waiting on the license before self-hosting Large 4 or fine-tuning it, read the license on Hugging Face first, especially if you plan to resell access to the model as an API. If you only need the model through an API, nothing changes in how you call it. To compare hosted costs with other models, use the AI token cost calculator.