KIMI K2.7 Code
Open SourceMoonshot AI•Released on 2026-06-12
Moonshot AI's coding-specialized open-weight model, appeared on Hugging Face June 12, 2026 under a Modified MIT license. A 1T-parameter native-multimodal MoE (32B active, 384 experts) with 256K context, built for long-horizon, end-to-end agentic software engineering. Company-reported results show it beating Kimi K2.6 on all six published coding/agentic evals (+21.8% on Kimi Code Bench v2, +31.5% on MLS Bench Lite) while using ~30% fewer thinking tokens. Scores 81.1% on MCPMark Verified, edging out Claude Opus 4.8's 76.4% on tool use. Independent third-party SWE-bench numbers were not yet available at launch.
Voice of the community
sample 40“Kimi K2.7 Code beats K2.6 on all six published coding and agentic evals, and uses roughly 30% fewer thinking tokens to do it.”
“K2.7 Code scores 81.1% on MCPMark Verified, edging out Claude Opus 4.8's 76.4% on tool use — a 1T coding agent shipping under a near-MIT license.”
Core Specs
Scenario Scores
Pros & Cons
Pros
- +Open-weight (Modified MIT) from day 1
- +MCPMark Verified 81.1%, beats Opus 4.8 on tool use
- +~30% fewer thinking tokens than K2.6 at higher scores
- +Native multimodal input (text + image + video)
- +Tuned for long-horizon agentic coding
Cons
- −No independent third-party SWE-bench numbers at launch
- −1T params require serious GPU infrastructure to self-host
- −Company-reported evals only; community data still forming