The Kimi-K3 open-weight model is the strongest on LisanBench, surpassing Gemini 3.1 Pro (high) but still trailing Opus 4.7 xhigh. It ranks 5th in the standard metric and 4th in a difficulty-weighted metric, despite using 2-3x more tokens than GPT-5.5 and Opus 4.8. Kimi-K3 exhibits a distinct search behavior, excessively using "wildcard families" where ~35% of its valid transitions involve changing only one letter in a word.