Kimi.ai’s announcement on X introduced Kimi-K2.7-Code as the company’s latest coding model, released with open-sourced weights and code. The post claims improvements over K2.6 of +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite, along with “30% lower reasoning-token usage” and better long-horizon coding performance. Kimi also says the model is available today through Kimi Code and the Kimi API, with a “6x High-Speed Mode” coming soon.
The company also launched a Kimi Code Beta Program for early access to upcoming models and features, and separately linked weights and code. In the launch graphics, K2.7 Code is compared with K2.6, GPT-5.5 (xhigh), and Opus 4.8 (xhigh) across internal coding and agent benchmarks. On the company’s own charts, K2.7 Code outscored K2.6 on every benchmark Kimi highlighted, though it did not lead the other frontier models on all tests. For example, it scored 62.0 on Kimi Code Bench v2, 53.6 on Program Bench, and 35.1 on MLS Bench Lite, while Opus 4.8 reached 67.4, 63.8, and 42.8 on those same benchmarks.
A second chart titled “Kimi-K2.7 Code vs Kimi-K2.6: Performance vs Tokens” places K2.7 Code at lower token usage than K2.6 on Kimi Code Bench v2, Program Bench, and MLS Bench Lite. The graphic is annotated with the note that “closer to the left (fewer tokens) is better,” reinforcing Kimi’s claim that the newer model is more efficient, at least on the company’s own measurements.
Kimi also published a pricing comparison card that lists K2.7 Code with a cache-hit rate of $0.19 per MTok, input at $0.95 per MTok, and output at $4.00 per MTok. K2.6 is shown at $0.16 cache-hit, $0.95 input, and $4.00 output, while K2.5 is listed lower still. That pricing card prompted some early questions in replies about cost, local running, and independent benchmarks, including calls for results on tests such as DeepSWE.
Source: Kimi Moonshot on X
