Mistral Large 4 weights are out: MIT with one clause, and that clause decides who can host it

Last week I wrote that Mistral Large 4's open weights meant downloadable bytes and nothing more until someone could read the license (the original Le Chonk serving piece). The bytes are now on Hugging Face, TechCrunch reported the release on October 29, and the LICENSE file is 2,632 bytes long. It answers most of my open questions, and it is friendlier than I expected for everyone except one group: companies that sell model access.

The license, clause by clause

It is called the Mistral Ecosystem Open-Weight License and describes itself as "a modified version of the MIT license". The grant is MIT's: use, copy, modify, merge, publish, distribute, sublicense and sell, plus explicit rights to deploy, further train and create derivatives. Condition one is the MIT attribution notice. Condition two is the new part: if the licensee "or any of its affiliates operates a Model as a Service business", it must sign a separate commercial license with Mistral "before using the Model or its derivative works for any commercial purpose".

The definition is narrower than it first sounds. Model as a Service means giving a third party access for inference or post-training, through an API or hosted endpoint, "in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data". Two things are carved out: end-user products where the model is "solely embedded within specific features or harnesses", and "mere relaying of requests to models hosted by others". Condition three exempts internal use, meaning anything that does not make the model available to third parties, and any use through Mistral's own products.

The last row is my reading of the word "affiliates", and it is the clause I would send to a lawyer. A conglomerate with a small GPU cloud subsidiary appears to trigger the condition for its commercial use even in an unrelated product. Compare that with the Kimi K3 license I used as a reference last time: no revenue threshold, no monthly-user branding clause, just a categorical line around one business model.

A ten-person startup selling ML4 endpoints needs a contract with Mistral first. A bank running it in-house for fifty thousand staff needs only the attribution notice.

The "meaningful control" test is the gray zone. A coding assistant where customers write free-form prompts arguably gives them control over inputs. I would not build a product on the harness exemption without written confirmation from Mistral.

What the released weights settle

The model card settles the parameter question I flagged, and both numbers I found earlier were right: 1T total, 49B active per token, and 52B once you count embeddings and output layers, hence the A52B in the repository name. In practice you store a trillion parameters and each token touches about 5% of them, because 4 of 128 routed experts fire per token, plus one shared expert. Add a 1M context, image input and text output, and the hard question stops being compute and becomes where a terabyte of weights lives.

Two checkpoints shipped. The FP8 one is 54 safetensors shards totalling about 1.06 TB, and Mistral's vLLM recipe serves it with tensor parallelism 8. The NVFP4 one keeps experts in NVFP4 and linears in FP8, comes to 32 shards and about 610 GB, targets B200, B300, GB200 and GB300, and needs an FP8 KV cache; Mistral's SGLang recipe runs it at tensor parallelism 4. By my arithmetic, 610 GB of weights across four GPUs only leaves useful KV room on B300-class memory, so on B200 I would still plan for eight. Both checkpoints embed a DSpark speculative decoding head, three draft layers proposing three tokens, which Mistral measured at up to 1.9x throughput.

What changed in the numbers and the price

The weights are a later checkpoint than the preview I wrote about. Mistral's release post cites three more weeks of reinforcement learning: Terminal-Bench 4.0 rose from 28% to 41%, and FrontierCode 2.0 is 19.7%. On the Artificial Analysis Intelligence Index v5 it scores 26, the highest among western open-weight models, with the leading Chinese open models still ahead on that aggregate. On cyber, Mistral reports a 96.6% refusal rate on malicious requests, and says live traffic showed potential uplift toward real attacks, so it is adjusting its guardrails. VP of Science Pierre Stock told TechCrunch that access "can't be cut off by definition because the model is open-weighted".

The API price halved: $0.70 input, $0.07 cached and $2.10 output per million tokens, against $1.36, $0.14 and $4.18 in my earlier piece. My reference workload of 2B input tokens a month, 80% cached, plus 200M output drops from about $1,604 to $812: 1.6B at $0.07 is $112, 0.4B at $0.70 is $280, and 0.2B at $2.10 is $420. That makes self-hosting for cost even harder to justify, since an 8-GPU Blackwell node rents for far more than $812 a month. The case for on-prem remains residency and control, which I covered in the sovereignty trade.

What I would do now

If you are an end-user organization, read the three conditions once, keep the attribution notice in your deployment, and self-host the NVFP4 checkpoint on Blackwell if residency is the requirement. If you build a product where customers type into the model, get Mistral's written view on the harness exemption before launch. If you run inference for others, contact Mistral before you put up an endpoint, because the license gives you no grace period and no revenue floor.