Mistral Large 4 Weights Are on Hugging Face, and the Card Disagrees With the Launch Post

Mistral AI has posted the weights for Mistral Large 4 on Hugging Face, ungated, under a custom licence that Hugging Face's metadata lists only as "other". The page was last modified on Oct. 29, according to the Hugging Face API, which is within the "end of this month" window Mistral gave when it opened a hosted preview.

The model card and the launch post do not match on several figures. As of Saturday afternoon, Mistral had not said which set is current.

What the model card says about Mistral Large 4

The model card describes a mixture-of-experts model of about 1 trillion parameters with 128 routed experts, 4 of them active per token, plus 1 shared expert. It takes text and images and returns text. The context window is 1 million tokens.

Mistral says the model was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European datacentres. The card gives no information about training data.

The card's examples assume 8 GPUs, and an NVFP4 checkpoint targets Blackwell hardware. It gives no memory figure.

Where the card and the launch post differ

The launch post, dated Oct. 6, says the model has "52 billion active parameters". The card's introduction says 49 billion active per token, and 52 billion only when embeddings and output layers are counted. The repository name carries the 52B figure.

The benchmark numbers differ as well. Both sets are Mistral's own.

Measure	Launch post (Oct. 6)	Model card
Active parameters	52 billion	49 billion (52 billion with embeddings)
Coding Agent Index	49.8%	48%
Terminal-Bench	28.3% (labelled 4.0)	41 (version not labelled)

The card attributes its 48% Coding Agent Index score to Artificial Analysis and says it is effectively tied with GLM-5.3 while using 47% fewer tokens per task. The card also gives an Artificial Analysis Intelligence Index v5 score of 26 and calls it the highest among western open-weight models. Mistral did not say whether the card or the launch post reflects a re-run.

Licence and access

The card calls the licence the Mistral Ecosystem Open-Weight License, "a modified version of the MIT license" for commercial and non-commercial use. Running a model-as-a-service business on it requires a separate commercial licence from Mistral. The launch post does not name a licence.

Mistral's hosted preview is priced at $0.70 per million input tokens and $2.10 per million output tokens, per the launch post. The model was ungated on Hugging Face as of Saturday.

Early uptake

Hugging Face showed 1,389 likes and 2,945 downloads for the repository on Saturday. A trillion-parameter checkpoint is too large for most individual users, so downloads are an unreliable guide to interest. Compare the figures on Nanbeige4.3-33B, a model small enough for one workstation, and on Liquid AI's LFM3, an earlier open-weight release this week.

Mistral's launch post also lists a Cybench score of 93% and a vulnerability reproduce-and-patch result of 82%. Both are vendor-reported, and the company said it was red-teaming the model with cybersecurity partners and public-sector bodies before release.

Mistral has not said when it will publish a corrected or reconciled set of figures.