Moonshot's open Kimi K2.7-Code gains 21.8% over K2.6 on its own code bench

Moonshot AI released Kimi K2.7-Code, an open-weight coding model under a Modified MIT license. The spec sheet is frontier-shaped: a 1-trillion-parameter Mixture-of-Experts that activates about 32 billion parameters per token (eight selected experts plus one shared, from 384 total), 61 layers, MLA attention with SwiGLU feed-forward, a MoonViT vision encoder adding 400M parameters for image and video input, native INT4 quantization, and a 256K-token context window. The weights, roughly 595GB on disk, are on Hugging Face for self-hosting via vLLM, SGLang, or KTransformers, alongside an OpenAI-compatible Kimi API. It is the model sibling of yesterday's Kimi Work, the local desktop agent that ran on the previous-generation K2.6.

The headline number is a 21.8% jump over K2.6 on Moonshot's own Kimi Code Bench v2, from 50.9 to 62.0, and the benchmark table around it is more candid than most vendor tables. K2.7-Code improves on K2.6 across all six benchmarks shown, but it still mostly trails the closed frontier: on that same headline bench it sits at 62.0 against GPT-5.5's 69.0 and Claude Opus 4.8's 67.4. Moonshot flags the one place it pulls ahead, MCP Mark Verified, a tool-use benchmark, where K2.7-Code scores 81.1 to Opus 4.8's 76.4. The model also reports roughly 30% lower reasoning-token usage than K2.6, framed as "less overthinking," which is the kind of efficiency that compounds in long agentic loops where reasoning tokens dominate the bill. And the price undercuts the frontier sharply: $0.19 per million cached input tokens, $0.95 on a cache miss, and $4.00 output, against Opus 4.8's $5 and $25.

The caveats are the usual ones for an open-weights release, and to its credit Moonshot mostly states them. Every headline figure is first-party, with independent verification still pending; the competitors were benchmarked in different environments while Kimi ran in its own CLI; and 595GB of weights means server-class hardware, not a laptop. Two operational quirks are worth knowing before you wire it in: thinking mode is mandatory and errors if you try to disable it, and the sampling parameters are locked server-side. For the thread we keep tracking, this is Chinese-lab open-weights pressure continuing down the coding lane: a frontier-shaped, MCP-native coding model at a fraction of the closed-model price, with weights you can actually host. It trails the very top, but the pitch is the one that keeps working, good enough, open, and cheap, and it quietly upgrades yesterday's story too, since Kimi Work's engine is no longer a reported K2.6 but an openly shipped K2.7-Code you can run yourself.