The model is an input.
Not a dependency.

A long list is easy. A list that costs you nothing is not.

Everywhere else, a model list means the software can call those APIs. It tells you nothing about the part you will actually feel. What happens to your standards the day you switch? The judgement was inside the model, and you have just replaced it.

Praxis keeps the judgement somewhere else. The thing that gates every step reaches the same verdict every time and has no model in it. The thing that audits that gate runs outside it and answers to the contract, not to the gate. Neither one ever asks what made the work.

So move down this list freely. Your cost changes. Your speed changes. Your evidence does not.

It also means the list is not a roadmap. A new model from a maker already here is one line of configuration. A new maker is one row. Neither of those is a release, and neither is us certifying anybody. There is nothing to certify when the checking never trusted the model in the first place.

What it runs on today.

Two kinds, and Praxis treats them exactly the same. Frontier models you reach over somebody else's API. Open weight models you can pull down and run on your own machines.

And whatever lands next. This page is built from the same catalogue the product reads, so it can only ever show what Praxis can actually be pointed at. Never a logo we wish we had.

Naming a model is easy. Getting the exact one is the work.

A roster is only worth something if Praxis can reach the model you are actually allowed to use. Sometimes that is a public download. Often it is not. So weights come from wherever yours already live, and the route is the same in every case.

The same name is not the same model.

Two runs of the same model can be two different things. A different quantisation, a different context window, a different runtime version, a different precision. Any one of those changes what the thing is capable of, and none of them change its name. A record that says only the name has not recorded very much.

So Praxis pins the exact release, checks the digest of every file it pulls, keeps track of where each file came from, and writes what served the work into the seal alongside the work itself. Ask later what produced a result and the answer is on the record rather than in somebody's memory.

Nobody can swap the substrate under a sealed engagement and have it pass quietly.

Which is also why moving down the roster is safe. The proof already says what it was standing on.

Or none of them, on hardware you own.

Every open weight model above can run inside your own network. On GPUs, on Gaudi, on TPUs, or on the accelerators already sitting in the machines you have. Nothing leaves the building.

And nothing gets left behind either. The whole apparatus travels with it. The Praxis pilot, a reasoning system built to judge rather than to generate, gating every step against the contract. The watcher that runs outside the loop and answers to the contract rather than to the pilot. The sealed record of every decision, written as the work happens. All of it speaks to a server in your own rack exactly the way it speaks to a hosted one, because the difference is an address rather than a different product.

That matters most where it stops being a preference. If your data cannot leave the country, or cannot leave the building, the model is no longer something you pick. It is a constraint. Praxis was built for that case first.