Short answers to what people ask about Mistral Large 4. Every number comes from the same sourced data as the rest of the site.
The weights alone need BF16 (2,100 GB), FP8 (1,050 GB) and 4-bit (525 GB). Memory for the context (the KV cache) and the inference software comes on top, so a setup that only just fits will struggle.
Sources: Mistral's model card
8× NVIDIA H100 SXM have 640 GB of memory together. That holds the weights at 4-bit (525 GB), but not at BF16 (2,100 GB) or FP8 (1,050 GB). Memory for the context (the KV cache) and the inference software comes on top, so a setup that only just fits will struggle.
Sources: NVIDIA's H100 page, Mistral's model card
8× NVIDIA H200 have 1,128 GB of memory together. That holds the weights at FP8 (1,050 GB) or 4-bit (525 GB), but not at BF16 (2,100 GB). Memory for the context (the KV cache) and the inference software comes on top, so a setup that only just fits will struggle.
Sources: NVIDIA's H200 page, Mistral's model card
Not on a single one. Apple's current Mac Studio specs list up to 256 GB of unified memory with the M3 Ultra, and even at 4-bit the weights alone need 525 GB.
Sources: Apple's Mac Studio specs, Mistral's model card
Mistral describes it as an open-weight model and says the weights will be released by the end of October 2026. The license has not been announced, so it is too early to say what you may do with them.
Sources: Mistral's model card, Mistral's announcement
$1.36 per million input tokens, $0.14 per million cached input tokens and $4.18 per million output tokens. The model card also lists a lower launch price ($0.68, $0.07 and $2.09) without saying when it ends.
Sources: Mistral's model card
1.05 trillion in total, of which 52 billion are active for each token, according to the model card. The announcement rounds the total to 1 trillion and says 49 billion are active.
Sources: Mistral's model card, Mistral's announcement
Mistral's model card says 1 million tokens. Artificial Analysis lists 524,288 tokens for the preview.
Sources: Mistral's model card, Artificial Analysis
Artificial Analysis, which runs the same tests on every model, gives the preview 38 on its Intelligence Index, where the median for comparable models is 26. In its own launch post, Mistral reports 61.7% on DeepSWE v1.1 and 93% on Cybench.
Sources: Artificial Analysis, Mistral's announcement
Yes. It takes text and images as input, with a 1.6 billion parameter vision encoder.
Sources: Mistral's model card
Mistral's launch post calls Mistral Large 4 “unofficially ML4, very officially: le Chonk” without saying why. In May, Mistral renamed its assistant Le Chat (“the cat”) to Vibe. In June, people on X joked about “Le Chaton Fat”, a giant Mistral model that does not exist, and Mistral's CEO posted “It's actually le gros chaton” (the big kitten).
Sources: Mistral's announcement, Mistral's Vibe launch post, a post by GLIF, Arthur Mensch's post
No. It is an unofficial fan site, not affiliated with or endorsed by Mistral AI.