At Catan, a small AI wins 36 out of 40 duels against large models

The VINCO model claims the top two spots in Catan duel on Colonist, with training that did not use any human games, according to its team.

Two accounts driven by the same AI took the top two global spots in the Colonist 1v1 Catan rankings, according to distributed cognition. The team places this achievement on November 1, after less than 24 hours of cumulative play. Its model, VINCO, is said to have about 500,000 parameters and was designed specifically for this competition.

The presentation report accompanies this announcement with screenshots of the rankings and a profile showing 107 games, with a 67.3% win rate. This record describes only one of the accounts presented. It should also be distinguished from the 90% win rate highlighted against large language models, obtained in another series of matches organized by the developers. A race to fifteen points, with no trading between players

The playing field matters as much as the ranking. In this variant, two opponents aim to reach fifteen victory points, instead of ten in the classic four-player version. Resource trades between players are disabled. The announced performance therefore concerns a strategic showdown, without the negotiation that plays an important role around a Catan table.

The rules of Colonist’s ranked mode also provide for balanced dice and a robber subject to early-game protection. The report describes 30-second turns, during which the agent must choose between building, buying cards, or keeping its resources. Opponents’ development cards remain hidden: deciding requires dealing with incomplete information while anticipating possible responses. Specialized training, with no human games

The announced method combines reinforcement learning on simulated games and self-play training. According to the team, no human game was used to train the model. The agent learns from game situations and the results of its actions, then refines its strategy by facing versions of itself.

The developers say they reached this level in less than a day of training on two RTX PRO 6000 Blackwell graphics cards. No total development cost is given in the report.

The team argues for an AI organization in which several specialized models would collaborate, rather than a single general-purpose model tasked with doing everything. Here, the game serves as a training ground for decision-making under uncertainty. The report presents the performance of one specialist; it does not test this collaboration between models. Forty matches to compare four large models

In the published evaluation, each opponent plays ten games against VINCO. The result shows nine wins for the specialist against Opus 5.5, nine against Fable 5.1, eight against Astra, and ten against Kimi K3. The total therefore reaches 36 wins out of 40, or 90%. The opponents were all set to a high reasoning level.

This sample remains limited: for each model, a single game changes the win rate by ten points. Nor is this a tournament in which all competitors face off against one another. The scores measure their success against the same specialist, without establishing an overall ranking among them.

A dedicated play platform is offered to challenge the agent. It uses its own ranking: VINCO’s rating there is set at 3,000, while those of the other AIs are estimated from the ten comparison matches. This value should not be confused with a rating obtained on the Colonist rankings. Earlier cities, less frequent ports

The gameplay choices described by the team offer another reading of the performance. The agent plays on average 53% more development cards than the highest-ranked players and favors the largest army over the longest road. It saves to build a city before expanding in 40% of cases, compared with 18% among the humans compared.

Later expansion may also benefit from the “friendly robber.” This rule protects players who have not yet reached three public points. The authors suggest that the agent lets the first 7 rolls pass before losing this protection; they present this explanation as their interpretation of its behavior.

Use of ports also differs from that of the human opponents observed. They use them in about 70% of their games, compared with half for VINCO. For specialized ports that allow two resources to be traded for one, the reported proportions are 47% for its opponents and 19% for the agent.