inferwire
/
AI·4 min read

Kimi K2.6 Wins One Round of an AI Coding Contest

In a sliding-tile word puzzle, bots written by ten language models played each other. Moonshot AI's open-weights Kimi K2.6 won, ahead of Xiaomi's MiMo V2-Pro and GPT-5.5.

TL;DR

  • Moonshot AI's open-weights Kimi K2.6 won day 12 of Rohana Rezel's AI Coding Contest, a word puzzle, ahead of MiMo V2-Pro and GPT-5.5 [^1].
  • It is one contest round, not a broad benchmark: the organizer says the top two were partly separated by seed variance [^1].

Background

Rohana Rezel runs the AI Coding Contest, which pits major language models against each other in real-time programming tasks with objective scoring [^1]. Each model writes a bot, and the bots then play a game against each other over a network connection to a game server [^1]. Kimi K2.6 comes from Moonshot AI, a Chinese startup founded in 2023, and its weights are publicly available [^1].

What happened

Day 12 was the Word Gem Puzzle, a sliding-tile letter grid between 10×10 and 30×30 tiles with one blank space. Bots slide tiles into the blank and claim English words that form in straight horizontal or vertical lines. Words of seven letters or more score their length minus six; shorter words cost points, and each word can be claimed only once. Each pair of models played five rounds, one per grid size, with ten seconds per round [^1]. Ten models entered, but code from Nvidia's Nemotron Super 3 had a syntax error and never connected, so nine competed [^1].

Kimi K2.6 won with 22 match points and a 7-1-0 record. Xiaomi's MiMo V2-Pro followed with 20, then GPT-5.5 with 16, Zhipu AI's GLM 5.1 with 15 and Claude Opus 4.7 with 12. Gemini Pro 3.1 and Grok Expert 4.2 scored 9 each, DeepSeek V4 scored 3 and Muse Spark 0 [^1].

The move logs show why. Kimi's bot was greedy: it made whichever slide unlocked the most valuable new words, and that volume paid off on the 30×30 grids, where the scramble had broken up nearly all the seeded words [^1]. MiMo's bot never slid at all and only claimed words of seven letters or more already on the board. Claude's bot didn't slide either and fell apart on the 30×30 grids, where tile movement was needed. DeepSeek's bot sent malformed data every round. Muse's bot claimed every word it found, short ones included, and finished at a cumulative −15,309 points [^1].

Why it matters

Rezel is careful about what the result shows. The puzzle tests real-time decisions and whether a model can write clean, working code that connects to a TCP server and plays a new game correctly; it doesn't test long-context reasoning or code generation from a spec [^1]. Kimi and MiMo finished two points apart with almost opposite strategies, which Rezel reads as partly seed variance. Rezel also lets a counterargument stand: heavily safety-tuned models may be more cautious about claiming words aggressively, so the scoring may favor some models' behavior over others [^1].

The wider point Rezel draws is about open weights. Kimi K2.6 scores 54 on the Artificial Analysis Intelligence Index, against 60 for GPT-5.5 and 57 for Claude, which Rezel calls "not parity, but it's close" [^1]. A model that anyone can download now lands near the top in contests like this one. The failures carry a lesson too: as Rezel notes, Muse read the task only partly and then executed that partial reading in full, which matters for anyone deploying models on structured tasks with penalties [^1].

Practical example

Say Tomás leads a small team that builds trading bots for a game studio's in-game market, where bad orders cost money. A colleague forwards the headline that Kimi K2.6 beat Claude, GPT-5.5 and Gemini, and asks whether the team should switch models.

Tomás reads past the headline. The contest measured one thing: a bot written once, playing a ten-second sliding-tile puzzle over a server connection [^1]. The winning margin was two points, which the organizer attributes partly to seed variance, and the ranking might look different under other scoring rules [^1]. That is not enough to swap the team's model.

The failures are more useful to Tomás. One model's code never compiled, one sent malformed data, and one ignored the penalty for short words and sank to −15,309 [^1]. Those are exactly the risks in a market with penalties for bad orders. So Tomás writes a test of the team's own: each candidate model writes a bot for a copy of the market, runs it against fixed seeds, and gets checked for code that runs, messages that follow the protocol and orders that respect the cost rules. Kimi K2.6 goes on the list because its weights are available to run locally [^1]. Whether it stays depends on that test, not on a word puzzle.

Related gear

We recommend this book because it covers how to evaluate foundation models for your own application, which is the step a single contest result can't replace.

AdvertisementAmazon

AI Engineering: Building Applications with Foundation Models

Sources

  1. [1]ThinkPol — An open-weights Chinese model just beat Claude, GPT-5.5, and Gemini in a programming challenge
  2. [2]Hacker News — Kimi K2.6 just beat Claude, GPT-5.5, and Gemini in a coding challenge