KOLIBRI-1 · RECORDED DEMO
Less talk. More Breakout.
Probabilities, not generated text.
Kolibri-1 reads the ball and paddle coordinates, scores four actions, and follows a paddle-tracking policy supplied in the prompt.
We used vLLM’s classification path to score action labels, without generating a text response. Screenshots are for spectators; the model receives structured state.
How the experiment worked
The scoring head reused selected rows of the model’s existing output weights. It was not a newly trained game-playing classifier. Scores are conditional on the supplied options and are not calibrated confidence. The model and controller both contribute to the gameplay shown.