Kimi K2.7 Code

2026-06-12 · Publisher : Moonshot AI · Domain : Artificial Intelligence

Moonshot AI releases Kimi K2.7 Code, an open coding-specialized model derived from K2.6, which cuts thinking tokens by about 30% while improving long-running software engineering tasks.

### Context Kimi K2.6 (April 2026) is a general-purpose model. For agentic coding, where the cost of thinking tokens weighs heavily, Moonshot AI releases in June 2026 a variant dedicated to code, on the same 1T-parameter base, and positions it against GPT-5.5 and Claude Opus 4.8. ### What the model brings - Derived from K2.6, oriented toward full software engineering workflows and long-running tasks. - About 30% fewer thinking tokens than K2.6 (vendor claim). - Gains on the benchmarks in the table: for example Kimi Code Bench v2 from 50.9 to 62.0, MCP Atlas from 69.4 to 76.0. - A “highspeed” version offered on the API, at about 180 tokens per second according to the platform documentation. - Thinking mode, text, image and video inputs. ### Architecture and specifications MoE with 1T parameters, 32B active, 61 layers (including 1 dense), 384 experts of which 8 are selected, 1 shared, MLA attention, 256K context, 160K vocabulary, MoonViT encoder (400M). License: modified MIT. Weights: huggingface.co/moonshotai/Kimi-K2.7-Code. Recommended engines: vLLM, SGLang, KTransformers. API identifiers: kimi-k2.7-code and kimi-k2.7-code-highspeed. ### Cost and access | Model | Input (cache miss) | Input (cache hit) | Output | |---|---|---|---| | kimi-k2.7-code | 0.95 | 0.19 | 4.00 | | kimi-k2.7-code-highspeed | 1.90 | 0.38 | 8.00 | Prices in USD per million tokens, official page consulted on 2026-09-29 (262,144-token context). Self-hosting hardware: not disclosed. ### Benchmarks Scores published by the vendor (Hugging Face card); Claude Code in xhigh mode for Opus 4.8 and Codex in xhigh mode for GPT-5.5, according to the notes on the card. | Benchmark | K2.6 | K2.7 Code | GPT-5.5 | Claude Opus 4.8 | |---|---|---|---|---| | Kimi Code Bench v2 | 50.9 | 62.0 | 69.0 | 67.4 | | Program Bench | 48.3 | 53.6 | 69.1 | 63.8 | | MLS Bench Lite | 26.7 | 35.1 | 35.5 | 42.8 | | Kimi Claw 24/7 Bench | 42.9 | 46.9 | 52.8 | 50.4 | | MCP Atlas | 69.4 | 76.0 | 79.4 | 81.3 | | MCP Mark Verified | 72.8 | 81.1 | 92.9 | 76.4 | Source: Hugging Face card of Kimi-K2.7-Code (vendor measurements). ### Versus closed models The vendor's table shows that K2.7 Code clearly improves on K2.6 but remains below GPT-5.5 on all six measures and below Claude Opus 4.8 on five of them. It surpasses Claude Opus 4.8 on MCP Mark Verified (81.1 versus 76.4) and comes close to GPT-5.5 on MLS Bench Lite (35.1 versus 35.5). The gap is larger on Program Bench (53.6 versus 69.1 and 63.8). The vendor therefore does not claim parity here, unlike K2.6, which measured itself against previous-generation models. The appeal remains the price (0.95 and 4.00 USD, official page) and the possibility of self-hosting. ### Key takeaways K2.7 Code is a specialized variant, a milestone in token efficiency rather than in level. It suits code agents where the cost of thinking matters, with a level still below the best closed models according to the vendor.

Tags : LLM, Open Source, Moonshot AI, Kimi, MoE, MIT modifiée, Codage