Kimi K2.7-Code Goes Open Source: Code Up 21%, Thinking Down 30%
Moonshot AI is back at it. Less than two months after K2.6, Kimi K2.7-Code is here — a model specifically polished for coding scenarios, fully open source. It improved across 6 major MCP and Code Bench benchmarks, putting more pressure on closed-source models.
💡 What You Will Learn
Moonshot AI is back at it. Less than two months after K2.6, Kimi K2.7-Code is here — a model specifically polished for coding scenarios, fully open source. It improved across 6 major MCP and Code Benc
📜 Table of Contents
Three Key Improvements
Kimi Code Bench v2: 50.9 → 62.0 (+21.8%) Program Bench: 48.3 → 53.6 (+11%) MLS Bench Lite: 26.7 → 35.1 (+31.5%) Kimi Code Bench is Moonshot's comprehensive code evaluation across multiple languages and scenarios. MLS Bench Lite focuses on ML engineering tasks — the 31.5% jump shows K2.7-Code's understanding of complex AI engineering has leveled up significantly. Agent capability gains: - MCP Mark Verified: 72.8 → 81.1 (+11.4%) - MCP Atlas: 69.4 → 76.0 (+9.5%) - Kimi Claw 24/7 Bench: 42.9 → 46.9 (+9.3%)
30% Less Thinking: The Real Win
The standout metric is hidden in the details. Reasoning token consumption is down ~30% compared to K2.6. K2.6 had a tendency to overthink — circling around simple problems. K2.7-Code is optimized to think less and answer more accurately. For practical use: if you run K2.7-Code in a coding agent (OpenClaw, Kimi Code CLI, etc.), each call uses 30% fewer reasoning tokens, meaning lower latency and lower costs. The cumulative benefit on long-horizon tasks (hundreds of tool calls) is dramatic. Architecture: 1T total params / 32B active MoE, 256K context, MLA attention, MoonViT visual encoder. Same family as K2.6, so existing deployment setups (vLLM/SGLang/KTransformers) work directly.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.
