Mistral Large 4's Open Weights Weigh 1.06 TB, and the Serving Recipe Starts at Eight GPUs

Mistral Large 4 is a downloadable model, but the hardware it implies is not casual. The Hugging Face repository is not gated and reports 1,060,619,283,859 bytes of storage, about 1.06 TB, while its serving examples use tensor parallelism across 8 GPUs.

The model card describes a mixture-of-experts system of about 1 trillion parameters with about 52 billion active per token. The number that matters for a buyer is the first one, because every expert has to sit in memory even though only a few run per token.

What the card says about size and serving

The architecture has 128 routed experts with 4 active per token, plus 1 shared expert. The card gives a 1M-token context window, text and image input, text output, and a configurable reasoning effort of low, medium or high. The intro cites 49B active parameters, or 52B including embedding and output layers.

Precision is the lever. The main checkpoint is tagged fp8, so roughly one byte per parameter, which is consistent with the 1.06 TB repository. A separate NVFP4 checkpoint targets Blackwell GPUs, and GGUF builds exist for llama.cpp.

The vLLM and SGLang examples both pass a tensor-parallel size of 8. The card lists no minimum VRAM, and it names official Docker images: vllm/vllm-openai:mistral-large-4 and lmsysorg/sglang:mistral-large-4.

Item	Figure	Source
Total parameters	About 1 trillion	Model card
Active per token	About 52 billion	Model card
Repository storage	1.06 TB	Hugging Face API
Tensor-parallel size in examples	8	Model card
Training hardware	3,800 Grace Blackwell GPUs	Model card, Mistral

The licence is not plain MIT

The card names the Mistral Ecosystem Open-Weight License, a modified MIT licence. Running a Model-as-a-Service business needs a separate commercial licence from Mistral. Hugging Face's metadata lists the licence simply as "other".

For an enterprise running the model internally, the terms are permissive. For a cloud that wants to resell it, they are not. That is the practical line between this release and an MIT-licensed checkpoint.

Vendor benchmarks, labelled as such

Mistral's October 6 announcement opened a public preview and said the weights would follow by the end of the month. It reports its own results: 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4.0 and 93% on Cybench. These are vendor-supplied. It also reports a score of 3.74 out of 5 in a Surge AI coding human evaluation, second of five models behind Claude Opus 5 at 4.22.

The card cites the Artificial Analysis Intelligence Index v5 at 26 and says that is the highest among Western open-weight models. That claim is Mistral's wording on its own card, not an Artificial Analysis publication that this reporting fetched.

Hosted pricing is $0.70 per million input tokens and $2.10 per million output tokens. Mistral does not say what an 8-GPU node costs to run, so no cost comparison with the API can be drawn.

Demand so far

The repository was created on September 30 and last modified on November 2. It shows 6,556 downloads and 1,525 likes. A download count for a 1.06 TB repository mostly reflects hubs, mirrors and quantisers, not people serving it.

Other large mixture-of-experts work faces the same constraint. The Terminal's reporting on Colibri, an engine that streams experts from disk and on FreeToken's 290B claim with no throughput are the nearest comparisons on how much memory such models demand; the figure all three still lack is delivered tokens per second at a given context length.

What would change this read

Two published numbers would settle it: tokens per second per node at a given batch size and context length on the fp8 and NVFP4 checkpoints, and a minimum-memory line on the card. Until then, an 8-GPU node is the only hardware Mistral has shown it running on.