Mistral Large 4's open weights are out - what the licence lets you do with your own automation

Mistral published the weights for Mistral Large 4 on 29 October, two days after the date it gave when the model went into preview. The checkpoint is on Hugging Face as `mistralai/Mistral-Large-4-1T-A52B`. According to the model card, that is roughly 1 trillion parameters with about 52 billion active per token, text and images in, text out, and a 1-million-token context window. Mistral's announcement lists the API at $0.70 per million input tokens, $0.07 cached and $2.10 output, starting with the release.

When we covered the preview on 6 October, we said to wait for three things before committing any infrastructure: the release date passing, the weights actually being downloadable, and a licence you had read. All three have now happened. This is the follow-up.

What changed since the preview

First, the price. Mistral's 6 October announcement gave a list price of $1.36 input and $4.18 output, with the preview at half. The release announcement lists $0.70 and $2.10 and names no end date. For comparison, Claude Sonnet 5.5 and GPT-6.1 Sol list at $2 input and $10 output, and GPT-6.1 Terra at $1 and $5.

Second, the scores, which are still Mistral's own. After three more weeks of training, Terminal-Bench 4.0 rose 13 points to 41%. On the Artificial Analysis Coding Agent Index it scores 48%, about level with GLM-5.3 while using 47% fewer tokens per task. On the B3 agent security benchmark Mistral reports 97.1% resistance to attacks, up from 93.3% in preview. That last figure matters more than it sounds for any agent that reads untrusted email or web pages. Mistral's own human-preference chart still puts it behind Claude Opus 5.5 on code. Strongest open model on several charts is a different claim from strongest model.

Third, the licence, which is the part we were waiting for.

What the licence says

It is called the Mistral Ecosystem Open-Weight License, a modified MIT licence. You can use the model, fine-tune it and sell products built on it. There is one condition. If you or any affiliate runs a "Model as a Service" business, you need a separate commercial licence from Mistral first.

The definition decides who that is. Under the licence text on Hugging Face, Model as a Service means giving a third party inference or training access, by API or hosted endpoint, in a way that lets them meaningfully control 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 simply passing requests on to a model someone else hosts. Internal use, meaning nothing is made available to outside parties, is exempt entirely. There is no revenue threshold.

What it means for a business owner

Read against the work most businesses want automated, this is a permissive licence.

An agent that triages your support inbox, pulls invoice fields into your accounting system, or drafts replies for staff to approve, running on servers you control, is internal use. No extra licence.

A customer-facing feature built on it, say a quote builder on your website or an assistant scoped to your own products, looks like the embedded-feature carve-out. Your customers use the feature. They do not get to set the model's parameters or send it whatever they like.

The grey area is when someone else hosts the model for you. If an agency or IT provider runs a Large 4 server and gives your team an endpoint to send arbitrary prompts to, that provider may be the one running a Model-as-a-Service business, and the commercial licence is theirs to hold. Ask them in writing. If the server sits in a cloud account you own, it is your internal use. That is one more reason to own the infrastructure as well as the code. This is a plain reading of the text, not legal advice.

The honest caveat

The hardware has not got any smaller. The FP8 weights are on the order of a terabyte before you serve a single request. Every serving example on the model card uses eight GPUs, and Hugging Face's own hosted option uses eight H200s. The 4-bit NVFP4 checkpoint roughly halves the memory but needs Nvidia's Blackwell cards. Community 2-bit versions run on a single Apple Silicon machine with 512 GB of memory, at a quality cost nobody has measured independently yet.

For most small and mid-sized businesses, then, the licence question stays theoretical. You will call the API, and at $0.70 and $2.10 that is a sensible thing to do. The weights act as insurance. If Mistral raises the price or retires the model, you or a hosting provider can run the exact same model and your automation keeps working. That is worth having. It is not the same as running it yourself today.

Every benchmark above is also self-reported. Independent results on real business workflows are still thin.

What to do about it

If you tested a workflow on the preview, as we suggested in October, rerun the same cases on the released model. The scores moved enough that your preview results are out of date. If data residency is why you are looking at Mistral, ask whoever would host it two questions before moving anything into production: whose account the server lives in, and whether they hold a Mistral commercial licence if their setup needs one.