Mistral Large 4 open weights are out: read the MaaS clause

Mistral published the Mistral Large 4 open weights on October 29, three weeks after the API preview we wrote about on October 6. The download is free and ungated on Hugging Face. The license is not Apache 2.0. It is a modified MIT license with one condition, and that condition decides whether your plan is free or a sales call.

What actually happened

From Mistral's announcement and the model card and LICENSE on Hugging Face:

• Size. About 1 trillion parameters, mixture-of-experts, 49B active per token (52B if you count embeddings and output layers). Text and image in, text out, 1M-token context.
• Better than the preview. Mistral says three more weeks of reinforcement learning went into these weights. Its own chart puts Terminal-Bench 4.0 at 41%, up from 28% in the preview. Those are vendor-reported numbers, and the Hugging Face blog lists different figures for some of the same tests.
• Checkpoints. FP8 and NVFP4. Every official serving recipe uses 8 GPUs (vLLM or SGLang, tensor parallel 8). NVFP4 targets Blackwell hardware.
• API price. Mistral now lists $0.70 per million input tokens and $2.10 per million output tokens, about half the preview price.
• The license. The Mistral Ecosystem Open-Weight License allows commercial use, fine-tuning, and redistribution. The exception: a business that operates "Model as a Service" must sign a separate commercial license with Mistral. The text defines that as giving a third party inference or post-training access with "meaningful control over the inputs, parameters, or training data." Internal use is exempt. So are end-user products where the model sits inside specific features.

Why it matters for your business

1. Internal use is clean. Run it for your own team's support drafts, contract review, or catalog work and the MaaS clause does not apply. There is no revenue cap and no user cap in the text.
2. A feature is not a service. A product description generator inside your app is "embedded within specific features." A raw chat endpoint that your clients point their own prompts at is closer to MaaS. If you resell AI to customers, design the product as a feature and get a lawyer to read section 2.
3. Self-hosting is still a rack decision. Eight datacenter GPUs is not a closet server. For most small teams the value of open weights is that OpenRouter, Microsoft Foundry, and other hosts can now serve it. More hosts means more price pressure and an exit if Mistral changes terms.
4. The API price fell by half. At $0.70/$2.10 it is worth a rerun of your eval set, even if you never download a byte.

Sources: Mistral AI, Hugging Face model card and license, Hugging Face blog.