Open-source Kolibri-1 Plays Breakout Zero-shot in 25ms per Decision

Developers showed that Aleph Alpha's open-source Kolibri-1 model can play Atari Breakout zero-shot, mapping action probability outputs to game controls without fine-tuning or text generation, at just 25ms per decision.

2026-10-06 ~ 2026-10-06 · 2 related posts

• Kolibri-1 plays Breakout with zero fine-tuning, 25ms per move — kmodi · 2026-10-06
• Open-weights Kolibri-1 plays Breakout with no fine-tuning at ~25ms per move — Nils_Reimers · 2026-10-06