Mistral's blog says 49 billion active parameters, its docs say 52 billion
Mistral published Mistral Large 4 on 6 October and gave it two different sizes on the same day. The announcement calls it "a 1 trillion-parameter natively multimodal model with 49 billion active parameters", while the model's documentation page, dated 6 October as well, says it "features 52B active parameters and 1.05T total parameters, and a 1.6B vision encoder".

A typewriter with a sheet of paper reading ARTIFICIAL INTELLIGENCE. Mistral says it trained Mistral Large 4 on 3,800 Nvidia Grace Blackwell GPUs in its own European datacentres.
Both pages still carried those figures when we read them at 22:42 UTC on 6 October, and neither corrects the other. Three billion active parameters and 50 billion total separate the two descriptions of a model Mistral is already serving to customers.
The active count is the one that says how much of the model runs. Mistral gives ML4 a "granular Mixture-of-Experts architecture", and in that design only a fraction of the parameters is switched on for any given token, which is what the active figure measures and the total figure does not.
Everything else on the two pages lines up
The documentation lists the model as mistral-large-4-0, with mistral-large-4 as an alias, a version string of v26.10, a maturity of "Public Preview" and a context window of 1M. Mistral's announcement points developers at the preview API on Mistral Studio.
On infrastructure there is no gap between them. ML4 "was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe", the announcement says, and "the public preview is served on that same infrastructure".
Mistral adds that a significant share of the training data was multilingual, "spanning more than 160 languages, including every official language of the European Union", and that it runs the European deployment "end-to-end, independently of other digital service providers and under European law". The company has a nickname for it too: "Unofficially ML4, very officially: le Chonk."
The totals could be a rounding apart, since 1.05T rounds to 1 trillion. The active figures cannot be, because 49 billion and 52 billion are both written out exactly, one in the announcement and one in the documentation.
A third number appears on only one of them. The documentation gives Mistral Large 4 a 1.6B vision encoder on top of the active and total counts, and the announcement, which calls the model natively multimodal, carries no encoder figure at all.
An open-weight model with no weights yet
Both pages describe ML4 as open-weight, and neither offers a download. "We will release the weights by the end of the month, along with more details on the architecture, additional benchmarks, and our post-training methodology", Mistral says, and until that happens the preview API is the only way to reach the model.
A published checkpoint carries its own parameter counts in the file, so the release Mistral has promised for the end of October would show which of the two figures describes what it built. Alibaba took a similar route in September with an image model published from open weights under a research-only licence, where the weights landed first and the terms decided what anyone could do with them.
Neither Mistral page has been edited to agree with the other, and the company has published no correction to either. For now the honest description of Mistral Large 4 is that its maker gives it 49 billion or 52 billion active parameters, depending on which of its own pages you open.