Kimi K2.7 Code
OPENCoding/agent-specialized variant built on the Kimi K2.6 base weights, tuned for long-horizon software-engineering workflows and thousands of sequential tool calls, with ~30% fewer thinking tokens than K2.6 for equivalent tasks. A same-weights 'HighSpeed' serving SKU (kimi-k2.7-code-highspeed) also shipped, at roughly 2x the price ($1.90/M input cache-miss, $8.00/M output) for ~180-260 tok/s output — this is a serving/throughput tier, not a reasoning-effort tier, so it is not modeled as a separate entry here. All benchmark improvements Moonshot has published for this model to date are on Moonshot's own proprietary/in-house suites (Kimi Code Bench v2, Program Bench, MLS Bench Lite, Kimi Claw 24/7 Bench, MCP Atlas, MCP Mark Verified); no standard public-benchmark numbers (SWE-bench, LiveCodeBench, AIME, GPQA) have been disclosed for K2.7 Code specifically, and no independent third-party re-run results were found as of this research.