Kimi K2.6 Wins Coding Challenge

In the Day 12 Word Gem Puzzle of the AI Coding Contest hosted by Rohana Rezel, Kimi K2.6 took first place with 22 points (7-1-0).

• MiMo V2-Pro came in second with 20 points, GPT-5.5 third with 16 points, and GLM 5.1 fourth with 15 points.
• Claude Opus 4.7 placed fifth with 12 points, while Gemini Pro 3.1 and Grok Expert 4.2 tied for sixth with 9 points each.
• DeepSeek V4 failed to make a valid play due to malformed output, and Muse Spark indiscriminately claimed short words, dropping to a cumulative -15,309 points according to the article.
• Nemotron Super 3 failed to even connect to the server due to syntax errors.

The puzzle involves claiming horizontal and vertical words on a 10x10 to 30x30 sliding tile board, with words under 7 characters incurring penalties. While static scanners that read the initial seed word could hold up on smaller boards, the 30x30 size required strategies that actually moved empty spaces to reconstruct words.

Kimi adopted a greedy sliding strategy, choosing next moves based on positive-scoring words newly revealed with each move, and the article notes that this approach yielded results on larger boards. Conversely, MiMo and Claude mostly relied on initial-grid-dependent strategies without sliding, while GPT-5.5 used limited sliding of approximately 120 moves per round.

The article does not overstate these results as a single benchmark but interprets them as a signal that open-weight models are rapidly closing the gap with frontier models.