Mistral's open weights exclude the host

On 29 October Mistral put the weights of Mistral Large 4 on Hugging Face, keeping the promise it made when the model previewed on 6 October. One trillion parameters, 49 billion active per token, a one-million-token context window, a quantised NVFP4 checkpoint for Blackwell GPUs alongside the full FP8 one. Europe's biggest model is downloadable. Read the licence before the benchmarks.

It is called the Mistral Ecosystem Open-Weight License, and Mistral describes it accurately as a modified MIT licence. The grant is generous: use, copy, modify, merge, publish, distribute, sublicense, sell, run, further train. Then section 2: if you or your affiliates operate a "Model as a Service" business, you need a separate commercial licence before any commercial use of the model or its derivatives.

Model as a Service is defined as giving a third party access to the model for inference or post-training "in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data". Two carve-outs: end-user products with model capabilities embedded in a feature or harness, and "mere relaying of requests to models hosted by others". Internal use is exempt entirely.

So you may embed ML4 in your own product, fine-tune it, run it inside your company, or pass your users' requests to Mistral's hosted endpoint. What you may not do is the one thing that would compete with Mistral itself: put the weights on your own GPUs and sell access. The licence does not just open the model. It fences off exactly the business Mistral is in.

The day before, Mistral became the thing it banned

On 28 October — one day before the weights dropped — Mistral announced its Responses API in public preview and said it is now serving DeepSeek V4.1 Flash and GLM 5.3 Flash on Mistral AI Cloud Inference in the EU and US. Mistral is a model-as-a-service operator for other labs' open weights. Under an ML4-style licence, that hosting would need a paid licence. DeepSeek and GLM use real open licences, so it doesn't. Mistral benefits from openness it has chosen not to extend to its own flagship.

I don't think that makes the release dishonest. It makes it precisely drafted. "Open weights" was already a weaker promise than open source; this licence makes the reservation explicit and names its boundary. You can argue about whether a modified MIT with a hosting clause still counts as open. You can't argue about what it forbids — it is written down, which is more than most labs give you.

Self-hosting means a datacenter, not a laptop

The other limit is physical. The FP8 checkpoint runs to roughly a terabyte of files; Mistral's own vLLM example shards it across eight GPUs with tensor parallelism. The quantised NVFP4 variant targets Blackwell hardware — reporting puts it on four B200 or B300 GPUs. "Run it yourself" here means owning or renting a node of datacenter-class cards. For nearly everyone, ML4 arrives as an API: Mistral's own, at a reported $0.70 per million input tokens and $2.10 per million output, or through Vibe, the coding agent powered by it.

The benchmarks are real and worth noting. Three extra weeks of reinforcement learning after the preview took Terminal-Bench from 28% to 41%, just ahead of GLM-5.3's 40%. On the Artificial Analysis Coding Agent Index, ML4 scores 48%, effectively tied with the leading open-weight model while using 47% fewer tokens per task — Mistral's own framing, citing an external index. On the Intelligence Index v5 it scores 26, the highest among western open-weight models; the card itself notes the leading Chinese models score higher.

The honest way to read this release is as two facts, not one. Europe now publishes frontier-class weights, trained on 3,800 Grace Blackwell GPUs in Mistral's own European datacentres, fluent in every official EU language. And Europe's frontier lab just defined openness as everything except the revenue line. Both are true. The first is the headline. The second is what the headline costs.

Sources: Mistral Large 4 model card and licence on Hugging Face, Mistral's 6 October announcement, and the licence text, all read 1 November 2026. The Responses API launch was announced 28 October; GPU and pricing figures as reported there and by forkalt's spec sheet.