Mistral's open-weight Large 4 comes with a catch for anyone reselling access

Mistral Large 4 is open, provided you are not selling access to it. That is the plain reading of the license attached to the model's weights, which Mistral published on Hugging Face on October 29, 2026. Mistral's pitch is "sovereign cyber defense capabilities to enterprises and governments globally." The license has a second clause that decides who actually gets that.

The clause that decides who can use it

The license lets anyone use, modify, fine-tune and sell the model. The condition is that a licensee or its affiliates that operate a Model as a Service business must sign a commercial license with Mistral first. The license defines that as giving a third party access for inference or post-training, through an API or hosted endpoint. The third party must be able to control inputs, parameters or training data.

Two exits are written in. Internal use is exempt, so a bank can run the model on its own servers without calling anyone. So is any use through Mistral's official products. A startup that wants to resell hosted access to its own customers has to negotiate.

Mistral's blog post calls the model open-weight. It reserves the words open source for the libraries it uses, such as NVIDIA's NeMo AutoModel and TensorRT LLM. TechCrunch, in its report on the release, used both terms for the model. Only one of them matches the license.

Benchmarks, mostly Mistral's own

On the Artificial Analysis Coding Agent Index, Mistral reports a score of 48 percent, effectively tied with GLM-5.3, the leading open-weight model, while using 47 percent fewer tokens per task. On the same firm's Intelligence Index v5 it scores 26, which Mistral describes as the highest among western open-weight models. The model card then adds that the leading open-weight models from China score higher on that aggregate.

The cyber numbers come from the blog post. It lists a 51 percent score on the Artificial Analysis Cyber Index and 41 percent on Terminal-Bench 4.0, a figure Mistral labels self-reported. It also claims a 96.6 percent average refusal rate on malicious cyber requests, which it calls the highest it measures among open models. That is Mistral choosing which models to measure itself against. The Register reviewed the October 6 preview and advised skepticism toward the company's charts. It placed the preview on Artificial Analysis's leaderboard between DeepSeek V4.1 Flash and OpenAI's entry-level GPT6 Luna.

A cyber model that cannot be recalled

The blog post contains an admission most launch posts would leave out. Mistral says it watched the model refuse black-box attacks against live targets, "but we see a potential for uplift to conduct real attacks," and that it is "modifying our guardrails accordingly." The post does not say how a guardrail change would reach copies that are already downloaded.

TechCrunch quoted Pierre Stock, Mistral's VP of Science, saying that open weights mean access "can't be cut off by definition because the model is open-weighted." He framed that as the point: companies can run the model on their own systems, and defenders get it as fast as attackers do. He also said Mistral does not want to fearmonger. Runtime Wire covered the staged rollout earlier in October. It noted that the control letting an organization set its own security policies can also make safeguards easier to strip out.

Mistral is not shipping its sharpest version to everyone. The blog says a specialized cybersecurity variant is in preview with Orange Cyberdefense, with access subject to rigorous vetting and contractual safeguards. The public weights are the general model, and the vetted tier stays behind a contract.

What running it takes

The weights are a 1 trillion parameter mixture-of-experts set, with a smaller NVFP4 checkpoint aimed at Nvidia's Blackwell GPUs. Mistral gives no official hardware requirements. The Register's estimate is that an 8-way GPU system such as Nvidia's HGX B300 or AMD's MI355X can run it reasonably well. That is a data center server, not a workstation. The first wave of users will be companies and governments, which is who Mistral says it built it for.

For anyone planning to build a product on it, the order of operations is simple. Read clause two of the license before the benchmark charts. Then ask whether a customer will ever send prompts to a server you operate.