Mistral Large 4 Weights Are Out: What It Takes to Run Them and What the License Allows

2026-10-07
Brian

· Updated
Subscribe
Get New Posts by Email
New RepoSquare posts in your inbox: comparisons, setup guides and explainers on open-source tools and AI coding agents. No spam, and you can unsubscribe anytime.

When Mistral Large 4 launched as a preview on 6 October 2026, the blog promised weights “end of this month”, and there was nothing to download. That’s changed. On 18 November 2026, Mistral’s Hugging Face account had two public, ungated Large 4 repos, and the docs card listed the model as generally available from 5 November under a license called the Mistral Ecosystem Open-Weight License.

So the question this post used to answer, “can I download it?”, now has a yes. The more useful question is whether you can run it, and what the license lets you do once you have.

What You Can Download Now

There are two checkpoints. The main repo, mistralai/Mistral-Large-4-1T-A52B, ships FP8 weights in 54 safetensors shards, and its config marks the quantization as FP8 with 128 by 128 blocks. The second, Mistral-Large-4-1T-A52B-NVFP4, is a 4-bit build in 32 shards that the model card says targets Blackwell GPUs and is supported by vLLM and SGLang. Both include the tokenizer, chat template and a technical report PDF.

The card describes a mixture-of-experts model with about 1 trillion total parameters and 49 billion active per token, or 52 billion counting embeddings and output layers. Each of its 72 layers routes to 4 of 128 experts plus one shared expert, and the context window is 1 million tokens. The docs card still rounds the total to 1.05 trillion with 52 billion active.

How Much Hardware Do the Weights Need?

More than most teams have. I summed the safetensors file sizes from the Hugging Face API and checked them against a few common eight-GPU node sizes, counting weights only with 10 percent headroom. Here’s the core of the script, followed by its full output, which also prints download times.

REPOS = {"FP8 (main repo)": 1060.6, "NVFP4": 610.0}          # GB, from the HF API
NODES = {"8 x 80 GB": 640, "8 x 141 GB": 1128, "8 x 192 GB": 1536}
for f, gb in REPOS.items():
    fits = [n for n, cap in NODES.items() if cap >= gb * 1.10]
    print(f, gb, fits or "none")
format            size GB  Oct guess  weights-only fit (10% headroom)
FP8 (main repo)     1060.6      1060  8 x 192 GB
NVFP4                610.0       627  8 x 141 GB, 8 x 192 GB

download time for the FP8 repo
     100 Mbit/s:  23.6 h
    1000 Mbit/s:   2.4 h
   10000 Mbit/s:   0.2 h

The FP8 build needs a node of eight 192 GB cards just to hold the weights. NVFP4 fits on eight 141 GB cards. Neither fits a node of eight 80 GB cards, and that’s before the KV cache for long prompts, which grows with context. The model card’s own vLLM example uses --tensor-parallel-size 8.

Comparison of the two Mistral Large 4 downloads: FP8 main repo with 54 shards, 1,060.6 GB, fits eight 192 GB cards, about 2.4 hours at 1 Gbit/s; NVFP4 repo with 32 shards, 610.0 GB, fits eight 141 GB cards, about 1.4 hours; both under the Mistral Ecosystem Open-Weight License

One small check worth reporting. In October this post estimated the file sizes by scaling Mistral’s Large 3 repos up to Large 4’s parameter count, and guessed about 1,060 GB for FP8 and about 627 GB for NVFP4. The real FP8 repo came in at 1,060.6 GB, and NVFP4 at 610.0 GB, a little smaller than the guess. Bytes per parameter carried over almost exactly.

For a sense of what a single consumer GPU can handle instead, running a 125B model on a 12 GB GPU shows where offloading stops being practical. A trillion parameters is far past that line.

What the Open-Weight License Allows

It’s a modified MIT license, 413 words long, and the change is one clause. You can use, copy, modify, fine-tune and redistribute the model and derivatives, commercially or not, as long as you keep the notice. But if you or an affiliate run a “Model as a Service” business, you need a separate commercial license from Mistral before using the model commercially.

The definition is what matters. Model as a Service means giving a third party access to the model for inference or post-training, for example through an API or hosted endpoint, in a way that lets them meaningfully control the inputs, parameters or training data. It excludes end-user products where the model is embedded in specific features, and it excludes relaying requests to models hosted by others. Internal use, where nothing is exposed to third parties, isn’t affected.

What you’re doingMy reading: commercial license needed?
Running it internally for your own teamNo
A product feature built on it, like a support botNo, if users can’t control the model directly
A public API that serves the model to customersYes
Fine-tuning and publishing a derivativeNo, keep the notice

It’s more open than the October docs suggested, and less open than Large 3, which shipped under Apache 2.0. If you’re a hosting company, read clause 2 before anything else. This isn’t legal advice.

Is the API Still the Practical Choice?

For almost everyone, yes. The docs card listed $0.70 per million input tokens, $0.07 cached and $2.10 output on 18 November. At those prices, a 3,000-token chat turn with an 800-token answer costs about $0.0038, and an agent run with 2 million input tokens, 1.6 million of them cached, plus 60,000 output tokens costs about $0.52. You’d need a lot of traffic to beat that with a rented eight-GPU node.

The benchmark picture hasn’t changed since launch. Artificial Analysis put the preview at 38 on its Intelligence Index, level with GPT-6 Luna at max effort and below the top of the chart, and Mistral’s “highest of any model” claim applies to one cyber subtest where some closed models scored near zero because they refused. A refusal score measures policy, not skill. Why OpenAI cancelled GPT-6.1 Astra covers how those refusals came about.

If you’re comparing API bills across vendors, Haiku 5.5’s 90 percent claim and what Gemini 4 Argon costs use the same per-call method. And Qwen4-Max is the opposite case, a big model you still can’t download at all.

The weights being public matters most for researchers, for companies that need everything on their own hardware, and for anyone building smaller fine-tunes. For a weekend project, it’s a 1 TB download you probably can’t load.

Get New Posts by Email
New RepoSquare posts in your inbox: comparisons, setup guides and explainers on open-source tools and AI coding agents. No spam, and you can unsubscribe anytime.

2 comments

Leave a comment