Distributed Cognition's 0.5M-parameter VINCO tops 1v1 Catan leaderboard and beats frontier LLMs
The lab says VINCO, trained with offline RL on simulated games and self-play and no human games, held first and second on Colonist.io on 1 November and won 36 of 40 games against Opus 5.5, Fable 5.1, Astra and Kimi K3; results are self-reported.
Why it matters
Catan mixes hidden information, dice randomness and adversarial play, a setting where the lab argues general LLMs plan poorly. VINCO reached the top of Colonist.io's 1v1 ladder in under 24 hours of play with about 0.5M parameters trained on two RTX PRO 6000 Blackwell GPUs, echoing how specialist RL systems beat generalists in Go and poker. The head-to-head numbers come from the lab itself, so independent matches would strengthen the claim.
Key facts
- About 0.5M parameters
- Trained on two RTX PRO 6000 Blackwell GPUs
- 1st and 2nd on Colonist.io 1v1 on 1 November
- Won 36 of 40 games against frontier LLMs