Kimi K3 在 Code Arena 前端基准上超越 Claude Fable 5，但复杂数学仍大幅落后

Moonshot's AI model Kimi K3 is getting a lot of attention in the Western AI community. The big question is how close it actually gets to the best Western models. Two new data points paint a mixed picture. In the Code Arena: Frontend benchmark, which ranks models based on human preference ratings, Kimi K3 scores 1,679, beating Claude Fable 5 (1,631), GPT-5.6 Sol (1,618), and every other tested model by a wide margin. It's the first time a Chinese model has claimed the top spot on this benchmark.

The picture looks different for hard math. According to data from Epoch AI, Kimi K3 hits only about 39 percent accuracy on FrontierMath Tier 4, the benchmark's hardest expert-level math tasks. Models from OpenAI and Anthropic score close to 90 percent there in some cases.