Moonshot Open-Sources Kimi K2.7-Code: A 1-Trillion-Parameter Coding Model, but the Benchmarks Come Only From Its Own

1 views

On June 12, Moonshot AI open-sourced the coding model Kimi K2.7-Code (1-trillion-parameter MoE, 256K context, Modified MIT license, weights on Hugging Face). The company claims 30% fewer inference tokens than K2.6, but all the eye-catching scores come from its own benchmark and haven't been verified by independent leaderboards.

Another Major Version

On June 12, Moonshot AI open-sourced the coding model Kimi K2.7-Code, the fifth major version in the Kimi series in under a year. It's a 1-trillion-parameter MoE (about 32B active, 384 experts), 256K context, Modified MIT license, with weights already on Hugging Face and also usable via the Kimi API and Kimi Code CLI. Pricing is about $0.95 per million input tokens and $4 output, positioned for "agentic coding," headlining long-chain tasks and token savings.

The Benchmarks Come Only From Its Own

The official figures are striking: on its own Kimi Code Bench v2, it's 21.8% higher than the previous generation K2.6, using about 30% fewer inference tokens. But one thing has to be spelled out—these numbers all come from Moonshot's own benchmark and haven't been submitted to independent leaderboards. Being open source and locally runnable are solid benefits; as for "how far ahead," that can currently only count as vendor self-attestation, so we recommend testing on your own real coding tasks rather than copying the percentages from the release notes.

A High-Speed Version on the Side

On June 15, Moonshot pushed a "HighSpeed" variant of Kimi Code Beta, claiming throughput of about 180 tokens/second for medium-length inputs and up to about 260 for short context—several times the standard version. Again, this throughput comes from the official notes, and the reality still depends on your own scenario.

via: Hugging Face (Moonshot AI open weights)