Kimi K3 Surpasses Claude Fable in Long-Duration Coding Benchmarks
Kimi K3 shows that Chinese open-weight models are now competitive with frontier US models on long-duration coding tasks, while still trailing in debugging.
Reporting from 1 source: ASCII.jp.
Moonshot AI announced the Kimi K3 model on July 16, with 2.8 trillion parameters and a 1 million token context. It outperformed Claude Fable 5 and GPT 5.6 Sol on the SWE Marathon long-duration coding benchmark, scoring 42.0. On Terminal Bench 2.1 it beat Claude models but fell short of GPT 5.6 Sol. The model is open-weight, with weights to be released by July 27.
Moonshot AI's Kimi K3 uses a Mixture of Experts architecture with 896 specialized experts, activating 16 per input. The company claims it is the world's first open model approaching 3 trillion parameters. Underlying technologies include Kimi Delta Attention for faster decoding in long contexts and Attention Residuals for training efficiency. The model also features native image understanding, enabling vision-in-the-loop for tasks like website design and game development. API pricing starts at $0.30 per million tokens for cached input.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.