Moonshot AI Ships Kimi K2.7-Code, an Open-Weight Coding Model It Says Uses 30% Fewer Reasoning Tokens

Moonshot AI has released Kimi K2.7-Code, an open-weight model built specifically for coding and agentic tasks, publishing the weights on Hugging Face on June 12 under a modified MIT license. It is the fifth release in the K2 line in under a year — following the original K2 base in July 2025, K2 Thinking in November, K2.5 in January, and K2.6 in April — and the first to be tuned as a coding-first variant rather than a general-purpose update.

The architecture is unchanged in headline terms: a Mixture-of-Experts design with 1 trillion total parameters and roughly 32 billion active per token, paired with a 256K-token context window. What Moonshot is selling this time is efficiency. The Beijing company says K2.7-Code cuts reasoning-token usage by about 30% versus K2.6, attacking the "overthinking" that inflates latency and API bills when a model talks itself in circles before producing an answer.

On its own evaluations, Moonshot reports gains of 21.8% on Kimi Code Bench v2, 11.0% on Program Bench, and 31.5% on the multi-language MLS Bench Lite over the previous version. Because the weights ship under a modified MIT license that permits commercial use with attribution, teams can self-host the model at no licensing cost and pay only for the hosted API if they would rather not run a trillion-parameter model themselves — a packaging that has made the K2 series a reference point for the open-weight camp.

Not everyone is convinced by the numbers. Independent testers note that the benchmarks are Moonshot’s own, and that "every model improves double digits on its own test suite." Researcher Elliot Arledge, who tested the release, called K2.7 "more honest but not more capable," reporting that it now writes its own GPU kernels instead of wrapping libraries — but that two of those kernels contained bugs and that kernel-optimization performance actually regressed against K2.6. Others pointed to K2.6’s modest 24% on the independent DeepSWE benchmark and asked whether the new model would submit to the same outside scrutiny.

The release lands as Chinese open-weight labs press their advantage on cost and openness while U.S. frontier labs lean on closed, premium models. It also arrives days after reports valued Moonshot at around $30 billion, a sign of how much capital is chasing the open-weight thesis. For developers, the practical takeaway is the one the skeptics keep repeating: the only benchmark that matters is your own workload, and a 30% efficiency claim is worth exactly what it saves on your bill once you measure it.