The model line.

We do not build models that talk their way to an answer. We build models that reason, and reason economically — more thinking per token, not more tokens per thought. Everything we have released so far sits on that single idea.

The scaling classes

The core of the catalog is one system scaled across three sizes. The classes are named for what they orbit: a satellite, a planet, a star.

Deimos R1 — Satellite class. The smallest of the line and our reasoning specialist. It carries the terse, token-efficient chain-of-thought that the whole family is built around, distilled down into a model you can run without ceremony.

Europa B1 — Planetary class. The medium model. Europa is where we ask how far intelligence scales per parameter, and where the reasoning techniques from Deimos get their first real room to breathe.

Ganymede A1 — Stellar class. The flagship. Ganymede takes the methods we refined below it and validates them where it counts — real tasks, real constraints, reasoning efficiency held as a first-class metric rather than a footnote.

The specialist

Not everything in the line scales the same axis. Kuiper R1 is a trace-inversion model: given another model's output, it reconstructs the reasoning that produced it. Recovering that chain-of-thought makes a result auditable instead of opaque — you can see the path, not just the destination.

Underneath

None of these were scaled into existence. They were distilled — the reasoning compressed, the traces curated, the token budget treated as something to spend deliberately. That method is its own story, and the next post is about the apparatus that makes it repeatable.