An open-weights Chinese model just beat Claude, GPT-5.5, and Gemini in a programming challenge

In a real-time AI coding contest, Moonshot AI's open-weights model Kimi K2.6 emerged victorious over major frontier models including Claude, GPT-5.5, Gemini, and Grok. The challenge: a sliding-tile puzzle where bots compete to find long English words under a 10-second clock. Kimi's aggressive sliding strategy allowed it to keep finding words when other models ran out of moves. Xiaomi's MiMo took second place without moving a single tile, demonstrating two opposite strategies achieving nearly identical scores. Claude and Grok also failed to slide, costing them on larger boards where reconstruction was essential.

Kimi K2.6 scores 54 on the Artificial Analysis Intelligence Index, closely trailing GPT-5.5 (60) and Claude (57). Crucially, Kimi's weights are publicly available — anyone can download and run it. This contest highlights that the once-formidable capability lead of closed frontier labs over open-weights models has become measurably small, a trend building for months. The result signals a shift in the AI landscape where open models can compete head-to-head with proprietary leaders in specialized tasks.