Mistral released the open weights of Mistral Large 4 in late October 2026, and the main repository on Hugging Face was last updated on 29 October 2026. Its files add up to about 1,060.6 GB. If your team is thinking of self-hosting it, budget for three jobs before you budget for a single GPU: read the licence, build an evaluation, and plan the hosting. Open weights are not a free API. They are a large file and a list of obligations.
What is Mistral Large 4?
Mistral Large 4 is Mistral's flagship model, a mixture of experts with about one trillion parameters that handles text and images. The model card on Hugging Face describes "a 1 trillion-parameter natively multimodal Mixture-of-Experts model with 49 billion active parameters per token (52 billion including embeddings and output layers)."
In its announcement on the Mistral site, the company says: "Today, we're releasing the weights of this improved Mistral Large 4." It points people to Hugging Face, and to vLLM or SGLang to serve the model on their own infrastructure.
How big is it to run?
Big. The file list of the main Hugging Face repository adds up to about 1,060.6 GB. That is before you load anything into memory, add a cache for long prompts, or keep a second copy for a rollback.
The model card gives an example serving command with a tensor-parallel size of eight, for "a node of 8 GPUs," and tells you to adjust to your setup. It also mentions an NVFP4 version that "targets Blackwell GPUs" with a smaller memory footprint. It does not give a minimum hardware requirement. Plan on a full multi-GPU node as the starting point, and test before you buy.
Only about 49 billion parameters are active per token, which keeps the compute per answer lower than a dense model of the same size. But all the experts still have to sit in memory. You pay for the whole trillion in hardware, even if each answer only uses a slice.
What does the licence actually say?
Mistral calls it the Mistral Ecosystem Open-Weight License, "a modified MIT license." I read the licence file in the repository, and the modification is specific.
- The grant is broad. You may use, copy, modify, fine-tune, distribute and sell the model, as with MIT.
- The condition is about hosting. If you or any affiliate "operates a Model as a Service business," you must sign a separate commercial licence with Mistral before using the model or its derivatives commercially.
- "Model as a Service" is defined. It means giving a third party access to the model for inference or post-training, for example 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 excluded. End-user products where model capabilities are "solely embedded within specific features," and simply relaying requests to models hosted by others.
- Internal use is exempt. The condition does not apply to use that does not make the model available to third parties.
So, in plain terms. Running Large 4 inside your company is fine. Putting it inside a feature of your product, where users cannot control the prompts or parameters, appears to be fine. Selling raw access to it, as an API that customers call with their own prompts, needs a deal with Mistral first.
The line between those last two will not always be clear. A product with a free-text box that sends user prompts straight to the model sits close to it. If your product does anything like that, ask a lawyer, and ask Mistral.
Why are open weights not a free API?
Because an API includes work you now have to do yourself.
When you call a hosted model, someone else buys the GPUs, keeps them running, patches the serving stack, handles traffic spikes and fixes it at night. Your bill covers all of that. When you download weights, you get the model and none of the rest.
What is the real work?
- Hosting. Hardware, a serving framework, autoscaling, monitoring and an on-call rota. A trillion-parameter model does not start in seconds, so cold starts and failover need thought.
- Evaluation. Vendor benchmarks are not your use case. Build a test set from your own tasks, with expected answers, and run it before and after every change. Without it you cannot tell if a new quantised version is worse.
- The licence. Decide which of your uses are internal, which are embedded features and which might count as Model as a Service. Write that down and keep it with the code.
- Security and data. Self-hosting keeps prompts on your servers, which is a real advantage for regulated data. It also makes you responsible for logs, access and retention.
When is self-hosting worth it?
When control matters more than convenience. If your data cannot leave your infrastructure, if you need to fine-tune deeply, or if you run enough volume that the hardware pays for itself, Large 4 is a serious option, and it is good to see a European lab ship weights at this scale.
If none of that applies, Mistral's hosted API is probably cheaper in practice, once you count the people.
What would I do this week?
Write down the one use case you would run on Large 4. Classify it under the licence: internal, embedded feature or Model as a Service. Then run your evaluation set on the hosted API first. If the model does the job there, you have earned the right to plan the hardware. If it does not, you have saved a terabyte of download and a lot of money.
Where Alongside fits
Deciding whether to host open weights or call an API, including the licence, is the kind of work covered by Alongside's technical advisory.