Moonshot AI has released Kimi K2.7 Code, an open-weight coding model under a modified-MIT licence, and the headline claim is restraint: it uses roughly 30% fewer thinking tokens than its predecessor K2.6 while scoring higher on the same benchmarks.
The architecture is a 1-trillion-parameter mixture-of-experts that activates only 32 billion parameters per token, with a 256K context window. On Moonshot's own Kimi Code Bench v2 it lifts from K2.6's 50.9 to 62.0. That still trails the closed frontier, with GPT-5.5 at 69.0 and Claude Opus 4.8 at 67.4 on the same test, but the gap on agentic coding is now single digits rather than a chasm.
The token cut matters more than the score. Long-horizon agent runs are billed by the token, and reasoning traces are where the meter spins fastest, so a model that thinks less to reach the same answer is directly cheaper to run in a loop.
The open question is whether Moonshot's in-house benchmarks hold up once independent harnesses get hold of the weights.