Moonshot AI's open-weights Kimi K2.6 topped a 10-model coding competition, outscoring GPT-5.5, Claude Opus 4.7, and Gemini on a complex sliding-tile puzzle challenge.
In Day 12 of an ongoing AI Coding Contest run by technologist Rohana Rezel, Kimi K2.6 — an open-weights model from Chinese startup Moonshot AI — won outright with 22 match points and a 7-1-0 record. Xiaomi's MiMo V2-Pro finished second, GPT-5.5 third, and Claude Opus 4.7 fifth. The task was a Word Gem Puzzle: a sliding-tile letter board requiring bots to form valid English words under time pressure, with scoring that rewards longer words and punishes short ones. All top-two finishers were open-weights models from Chinese labs, with every Western frontier lab finishing below them.
Kimi K2.6 is open-weights, meaning you can self-host or access it via API without paying OpenAI or Anthropic rates. Its win on a constrained search-and-optimization task (sliding tiles, time limits, word scoring) suggests it handles algorithmic reasoning and competitive programming under time pressure — not just chat. If your stack leans on GPT-5.5 or Claude for code generation or agentic loops, this is a credible alternative to benchmark against your actual workload.
Pull Kimi K2.6 via the Moonshot AI API or a compatible open-weights host and run it against your existing GPT-5.5 or Claude prompt suite for your highest-cost code generation task — measure token cost vs. output correctness on real inputs, not synthetic benchmarks.
Install the openai-compatible client: pip install openai and get a Moonshot API key at platform.moonshot.cn
Tags