Mistral Large 4 Weights on Hugging Face: Release Date & License

Weights status as of Oct 6, 2026

The repo exists, but the model files do not yet. Mistral’s docs tag Mistral Large 4 as “Open”, while the Weights tab on the model card and the Hugging Face page both show the files as not yet released.

When will Mistral Large 4 be on Hugging Face?

The safest date is Oct 31, 2026. Mistral’s announcement promises the weights “by the end of the month”, and the Hugging Face repo shows a current ETA of Oct 31. Several outlets, including The Next Web and Forkast, plus a models.dev pull request, give Oct 27 instead; no Mistral page confirms that earlier date.

Before the public release, Mistral is giving cyber leaders, vetted partners and state authorities a build with reduced moderation for red-teaming, so security teams can test it before the files go public.

What the repo name tells you

The repo name Mistral-Large-4.0-1T05-A52B encodes the model’s size. “1T05” means 1.05 trillion total parameters, and “A52B” means 52 billion active parameters per token. Mistral’s blog says 49B active; the Hugging Face page explains that 52B counts the embedding and output layers and 49B does not.

How big will the download be?

Expect roughly 1 TB or more. Mistral has not said which precision it will publish, so the table computes the weight files alone at the three common formats (parameters × bytes per parameter). KV cache, activations and runtime overhead come on top.

For comparison, Mistral Large 3 shipped as FP8 weights that Mistral said run on a single node of B200s or H200s. If Large 4 follows the same pattern, an 8× H200 or 8× B200 server is the realistic starting point; the run-locally guide has the full hardware math.

Mistral Large 4 license

Mistral has not announced the license for Mistral Large 4 yet; the Hugging Face page shows none and the docs only tag the model as “Open”. For Mistral Large 3, the license was published in the Hugging Face repo together with the weights.

The previous generation is the best reference. Mistral Large 3 (675B total, 41B active, released Dec 2, 2025) shipped under Apache 2.0, which allows commercial use and modification. Whether Large 4 uses the same terms decides if you can fine-tune and sell products on it without a separate agreement, so read the license file on release day before you build on the weights.

How to download it once the files are live

The repo ID is already fixed, so the standard Hugging Face CLI command will work as soon as Mistral publishes the files. Check free disk space first: at FP8 the weights alone are about 1.05 TB.

pip install -U huggingface_hub
# log in if the repo asks you to accept terms
hf auth login
hf download mistralai/Mistral-Large-4.0-1T05-A52B --local-dir ./Mistral-Large-4

What to check on release day

The first hours after the files appear answer the open questions about self-hosting.

• License file: Apache 2.0 like Large 3, or a custom license with commercial limits.
• Precision of the official files (BF16, FP8 or NVFP4) and the total download size.
• The recommended serving stack and minimum versions, as Mistral listed vLLM 1.12.0 and mistral_common 1.8.6 for Large 3.
• Recommended sampling settings and context length for self-hosted use (the API serves 524,288 tokens).
• Community quantizations (GGUF, 4-bit) appearing under the repo’s model tree.

Use Mistral Large 4 before the weights arrive

You do not need the weights to try the model. Our homepage chat runs Mistral Large 4 with 3 free messages a day and no account. Developers can call it today through the Mistral API at $0.68 / $2.09 per 1M tokens on the current sale (see the API guide and API pricing pages), and the updates page tracks the release as it happens.