You can download the AI. Can you afford to run it?

The download button is the easy part.

Mistral has released the weights for Large 4, giving people access to the learned parameters needed to run its model outside the company’s hosted service. The official repository records an initial release on October 29. That is a real change from the October 6 API preview, which promised downloadable weights later in the month. Release history.

Then you look at the files: roughly 1.06 terabytes in the main repository. Suddenly, “run your own AI” sounds less like installing an app and more like an infrastructure decision. Official weight files.

That gap is the story. Open access can move power away from a provider. It does not make the machinery disappear.

The bill starts after the download

Mistral describes Large 4 as a trillion-parameter mixture-of-experts model, with about 52 billion parameters active per token including embeddings and output layers. Its published serving examples use multiple GPUs. The smaller active count does not mean the rest of the model vanishes from the deployment problem. Model card.

The repository’s 1.06 TB figure is file size, not a measured RAM requirement, a minimum hardware specification or a monthly bill. Those depend on how the model is stored and served, how much context it handles, and how many people use it at once.

There is also a smaller NVFP4 version. Mistral says it reduces the memory footprint and targets NVIDIA Blackwell GPUs. That matters: it would be misleading to present the largest checkpoint as the only option. It would be equally misleading to turn a compressed checkpoint into a promise of cheap, effortless laptop use.

We have not downloaded or run either version, priced a deployment, or independently reproduced the performance claims.

Open weights, with conditions

The license deserves its own look before anyone builds a business around the download.

The Mistral Ecosystem Open-Weight License permits broad use and modification, but adds conditions to the MIT-style terms. If a licensee or an affiliate operates a Model-as-a-Service business, it requires a separate commercial license before commercial use, subject to stated exceptions.

The text exempts internal use and use through Mistral’s official products. Its definition of Model-as-a-Service also excludes certain products with model capabilities embedded in specific features and services that merely relay requests to other hosts. This is not a blanket ban on commercial applications. Nor should buyers assume it is unrestricted MIT software. Read the actual terms for the intended business.

Who gets the keys?

For an organization already equipped to run it, the upside is meaningful: more choice over where a model operates, how it is integrated and when its deployment changes. Mistral’s stated ambition is to give customers more control through self-deployment. Company announcement.

Our interpretation is that this could reduce dependence on a single hosted service. It also shifts responsibility onto the operator. Someone still has to maintain the system, protect access, test its answers and decide where human review belongs. Downloadability alone establishes none of those safeguards.

For smaller teams, the practical route may still be a hosted service or a smaller model. Owning the deployment is valuable only if that control solves a problem worth its cost.

Mistral has opened another door to advanced AI. The interesting question now is how many organizations can afford to walk through it, and what they will build when they do.

AI illustration. Bright reviewed the public release artifacts and documentation; it did not test the model.