Ten models participated in the "Word Gem Puzzle" held on the 12th day of the AI Coding Contest. According to [Hacker News], Kimi K2.6, an open-weight model provided by the Chinese startup Moonshot AI, won the competition with 22 match points (7 wins, 1 draw, 0 losses).
Xiaomi's MiMo V2-Pro took second place, followed by GPT-5.5 in third, Zhipu AI's GLM 5.1 in fourth, and Claude Opus 4.7 in fifth. Models from major Western labs all finished outside the top two.
The challenge is a puzzle where players slide letter tiles and a single blank space on a grid to form English words. Points are awarded for words of seven letters or more, while requesting shorter words results in a penalty. Each pair competed over five rounds across five different grid sizes, with a time limit of 10 seconds per round.
Kimi K2.6 earned points using a greedy strategy that prioritized moves to unlock scorable words. This approach was reportedly particularly effective in large 30x30 grids that required tile restructuring. It recorded a cumulative score of 77, the highest in the competition.
Meanwhile, models such as the second-place MiMo V2-Pro and Claude adopted a strategy of not sliding the tiles. While this allowed them to score quickly on grids where words were already present in the initial state, it is reported that they reached their limit and failed to score on large grids where sliding was necessary. Additionally, Nvidia's Nemotron Super 3 was unable to connect to the server due to syntax errors, meaning the competition was effectively contested by nine models.
Source: Kimi K2.6 just beat Claude, GPT-5.5, and Gemini in a coding challenge（HN 380pt・219コメント） (HN Search (backfill), 2026-05-03)