The competition’s format was creative: competing models needed to form English words in a letter grid ranging from 10×10 to 30×30 by sliding adjacent tiles. Words under seven letters incurred point penalties, while words of seven letters or more scored based on length. Each pair of models played five rounds with a ten-second time limit per round.

Ten models competed, and the results surprised many observers:

• 1st Place: Kimi K2.6 (Moonshot AI) — 22 points, 7-1-0
• 2nd Place: MiMo V2-Pro (Xiaomi)
• 3rd Place: GPT-5.5
• 5th Place: Claude Opus 4.7

No model from Western frontier labs made the top two.

Kimi K2.6’s strategy was aggressive sliding — it employed a greedy algorithm that at each step sought the move unlocking the most new words. On smaller grids, this approach was less effective, but on the 30×30 large grids, when other models lost competitiveness due to their inability to slide tiles, Kimi K2.6’s continuous output advantage became decisive.

Notably, Kimi K2.6 is an open-weights model — anyone can download and use it. It currently scores 54 on the Artificial Analysis Intelligence Index, compared to 60 for GPT-5.5 and 57 for Claude. While not yet at parity, the gap has narrowed considerably — and this comes from a freely available model.

This result marks an important shift in the AI capability landscape: a year ago, the prevailing assumption was that Western frontier labs held an insurmountable lead in model capabilities that open-source models could not close. Today, that gap is rapidly shrinking.

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