Distributed Cognition 推出 VINCO：0.5M 参数模型在 Catan 中战胜 Claude
一个 0.5M 参数的小模型在卡坦岛赢了 40 局里的 36 局，对手是 Claude Opus 5.5 和 Kimi K3，体积还小了 1000 万倍，思路很反直觉。
Distributed Cognition 发布 VINCO，一个仅 0.5M 参数的小模型。在 40 局卡坦岛（Catan）对局中，它赢了 36 局，击败对手包括 Claude Opus 5.5、Claude Fable 5.1、Astra 和 Kimi K3。按该公司统计，VINCO 的参数量比这些前沿 LLM 小约 1000 万倍。卡坦岛考验的是在不完全信息下做决策，这是编码基准测不出的能力。公司认为超级智能会来自多个这类专用模型协作，而非单一全能模型。
Distributed Cognition’s VINCO, a 0.5M-parameter model beat Claude Opus 5.5, Claude Fable 5.1, Astra and Kimi K3 in 36 of 40 in the the board game Catan.
Catan tests a skill coding benchmarks don’t: picking good moves without seeing the opponent’s cards or knowing the next dice roll.
The company’s broader view is that superintelligence will come from many models like this collaborating, rather than one model covering everything. By its numbers, VINCO is about 10 million times smaller than the frontier LLMs it beat in testing.
