Kimi K3.1 Pro Becomes First Open-Weight Model to Score on MazeBench

Kimi K3.1 Pro scored 2% on the MazeBench spatial reasoning benchmark, becoming the first open-weight model with a non-trivial score, ahead of DeepSeek v4.1 Flash at 1%. It solved a multi-step maze puzzle on its 20th attempt, a feat only five models have achieved.

2026-10-25 ~ 2026-10-25 · 3 related posts