Can we bring our own models?

Yes, provided the models are open-weight. Substrate is built to serve open models of several kinds, and a site that already relies on particular open language, vision, speech, embedding, or vision-language-action models can bring them onto the appliance. The constraint is not whose model it is but whether its weights can be loaded onto the Czar. If they can, Substrate serves it. If the model exists only behind someone else's hosted API, it cannot, because the appliance has no path to that API and nothing to load.

What bringing a model looks like

A model arrives on the appliance the same way a software update does: as a signed artifact that the customer brings in. On a connected site it is moved onto the facility network from the customer's own machine. On an air-gapped site it is carried in on media. The appliance verifies what it is given before loading it.

Once loaded, the model is served by Inference through the one API, alongside whatever else is resident on the pool. Callers address it the way they address any model, declare their latency budgets, and get answers or fall back. Evaluation records every call to it as it does for every other model. From the floor's side there is no difference between a model already on the appliance and one the site brought.

The same is true for models a robot vendor supplies with its fleet. If the vendor provides open weights, or an adapter to an open base model, the customer loads it and the vendor's robots call it on the pool. The vendor's model runs on the customer's appliance, and nothing about it leaves the building.

What a site can do with it afterward

A brought-in model is a starting point. Adaptation can fine-tune it on the site's own captures and documents, entirely on the appliance, producing an adapter that reflects this floor rather than the general case. The adapter belongs to the customer, stays on site, and is served alongside the base model it modifies.

This is often where the value is. A vision model that arrives knowing labels in general becomes one that knows the labels this site's suppliers actually print. A language model that arrives knowing maintenance language in general becomes one that knows how this site's technicians describe a fault. The site's data does the work, and the site keeps the result.

A site can also run several brought-in models at once, swap one for another, and compare them through the traces Evaluation keeps. If a new model behaves worse on the floor than the one it replaced, the site goes back to the previous one. The record shows which model answered each call, so the comparison is concrete.

What the site does not have to do

The site does not have to ask permission, send the model anywhere, or register it with a service. It does not have to settle for a model someone else picked. It does not have to change the applications or robots that call the API, because the API is stable and hardware-independent and the new model sits behind the same interface. The platform overview at /platform describes the model kinds Substrate serves, and the only question a site has to answer before bringing a model is whether the weights are theirs to load.