On June 12, Dark Side of the Moon Kimi released and open sourced the Kimi K2.7 Code programming model. Officials stated that internal and external benchmark evaluations show that compared with the K2.6 model, Kimi K2.7 Code has significantly improved the instruction following ability in long-context programming scenarios and the performance of long-range programming tasks. It has also greatly improved the tendency to overthink in long-range tasks, and the average token consumption has been reduced by 30%.
In internal and external benchmark tests to evaluate code capabilities, the performance of K2.7 Code has been significantly improved compared to K2.6: Kimi Code Bench v2 has improved by 21.8%, Program-Bench has improved by 11%, and MLS Bench Lite has improved by 31.5%.
In the Kimi Claw 24/7 Bench, MCP Atlas and MCP Mark Verified benchmark tests that evaluate the agent's autonomous execution capabilities, the performance increased by about 10%.
K2.7 Code provides services through the Kimi API open platform and Kimi Code Plan from now on. The standard input and output prices of 1M token are consistent with K2.6 (6.5 yuan and 27 yuan), and the cache input price is 1.3 yuan. In addition, a high-speed version of K2.7 Code will be launched on June 15. The output speed is about 5-6 times that of the ordinary version (about 180 token/s for regular scenarios and 260 token/s for short context), and the price is twice that of the ordinary version.
