yesterday, moonshot ai officially launched its new large-scale programming-specific model, kimi k2.7 code, and simultaneously open-sourced it. this model focuses on core development scenarios such as code generation, comprehension, and debugging, with deep optimizations for long-context instruction execution, modeling complex logical workflows, and resource efficiency. according to official disclosures, in multiple authoritative programming benchmarks, k2.7 code demonstrates significant improvements over k2.6: kimi code bench v2 shows a 21.8% gain, program-bench achieves 11%, and mls bench lite reaches an impressive 31.5%; moreover, its autonomous task orchestration capabilities improved by approximately 10%. crucially, the model’s tendency toward excessive reasoning has been substantially mitigated, reducing average token consumption by 30% and delivering more precise, efficient responses.
effective immediately, developers and enterprise users can directly invoke k2.7 code through the kimi api platform. pricing remains consistent with the previous standard—input/output costs of 6.5 yuan and 27 yuan per million tokens, respectively—with the cache-hit input rate reduced to 1.3 yuan. the kimi code plan service now defaults to k2.7 code. please note that this model is specialized for programming tasks; for non-code requests, we recommend continuing to use the more versatile k2.6. additionally, thinking mode must be enabled; otherwise, the api will return an error, and the kimi code interface will automatically downgrade to k2.6.
furthermore, moonshot ai announced that the high-speed version of k2.7 code will be available on the kimi api platform starting june 15. leveraging architecture-level acceleration, this version delivers output speeds of up to 180 tokens per second under typical programming workloads, with peak rates exceeding 260 tokens per second in short-context scenarios—roughly 5–6 times faster than the standard edition. meanwhile, pricing has only increased to twice the standard rate, offering exceptional cost-effectiveness while unlocking unparalleled coding efficiency.
