Everything we’ve released so far comes from one idea: a model should reach the right answer without talking itself in circles. We want more thinking in every token, and fewer tokens spent getting there.
The scaling classes
Three of the models are the same system at three sizes. The classes are named for what they orbit: a satellite, a planet and a star.
Deimos R1 (Satellite class). The smallest of the line, and our reasoning specialist. It carries the short, token-efficient chain of thought that the whole family is built around, in a model you can run without fuss.
Europa B1 (Planetary class). The medium-sized model. With Europa we ask how far intelligence scales per parameter, and it’s where the techniques from Deimos first get room to stretch.
Ganymede A1 (Stellar class). The flagship. Ganymede takes what we refined on the smaller models and tests it on real tasks with real constraints, where efficient reasoning matters most.
The specialist
Not everything in the line scales the same way. Kuiper R1 is a trace-inversion model: give it another model’s output and it reconstructs the reasoning that produced it. Getting the chain of thought back makes a result easier to audit, because you can see the path as well as the answer.
Underneath
None of these models were scaled into existence. We distilled them: compressed the reasoning, curated the traces, and treated the token budget as something to spend on purpose. The tools that make that repeatable are the subject of the next post.
