Kimi K2.7 Code (Moonshot AI)

Kimi K2.7 Code, last tested on 2026-08-25, is categorised as a Large Language Model for coding, provided by Moonshot AI, with 1T total parameters, 32B activated per token (MoE), a 256K-token context window (262,144), and a Modified MIT license for open weights. It is positioned as a coding-focused agentic model built u The snapshot lists API pricing at $0.95 per 1M input tokens and $4.00 per 1M output tokens via the Kimi API, with open weights available for self-hosting at no cost, and platforms including API, HuggingFace, Ollama, vLLM, SGLang, and KTransformers. Architecture details include Mixture-of-Experts with Multi-head Latent

Kimi K2.7 Code is Moonshot AI's coding-specialised model built on top of Kimi K2.6, using a Mixture-of-Experts architecture with 1 trillion total parameters, 32 billion activated per token, 384 experts (8 selected plus 1 shared), a 256K-token context window, and native INT4 quantization. It ships in two flavours — kimi The review tabulates Moonshot's published six benchmark comparisons against K2.6, GPT-5.5, and Claude Opus 4.8, with Kimi Code Bench v2 scores of 50.9 (K2.6), 62.0 (K2.7 Code), 69.0 (GPT-5.5), and 67.4 (Opus 4.8). It also covers pricing, the reasoning-token paradox between Moonshot's claims and user reports, pros and c

Future Stack Reviews' research briefing covers GLM-5.2 and Kimi K2.7 Code as large open-weight coding models released in the same June 2026 window, with permissive licenses and API prices far below the US frontier. The post notes that Kimi K2.7 Code ships natively quantized to 4-bit and still requires approximately 640 Its verdict advises using GLM-5.2 or Kimi K2.7 Code for non-sensitive coding work after independent evaluation, and warns that the "sovereignty" framing is conditional: a license grants permission, not operation. Developers who cannot self-host the build they rely on and verify the data path are making an API governanc

ThursdAI's June 2026 release roundup lists Kimi K2.7 Code among 32 AI releases tracked that month, explicitly attributing it to Moonshot AI with "1T MoE parameters" and the label "open-sources Kimi K2.7 Code for agentic coding." The model is grouped as a top release alongside Aria, GLM-5.2, Gemma 4 12B, and MAI-Thinkin The roundup confirms the June 2026 launch window and open-sourcing of the 1-trillion-parameter mixture-of-experts coding model from Moonshot AI, situated within a broader monthly wave that also included GLM-5.2 (753B open MoE) and Nemotron 3 Ultra (550B). Eight open-weights models shipped in June 2026, with Kimi K2.7 C

On June 12, 2026, Moonshot AI released Kimi K2.7-Code, a 1T-parameter Mixture-of-Experts coding model with 32B active parameters, 384 experts, and a 256K context window, open-weighted under a Modified MIT license on Hugging Face. The release is positioned as a coding-specialist execution refresh of the K2 MoE backbone Thinking mode is required, with non-thinking requests falling back to K2.6. The guide frames K2.7-Code as best for long-horizon agent coding and MCP workflows, and not suited for general chat. Subsequent updates appended to the article note Kimi K2.7 Code becoming generally available in GitHub Copilot on July 3, 2026 a

Puter.js now supports Kimi K2.7 Code, Moonshot AI's open-weight coding-agent model built for long-horizon, autonomous software engineering, as announced on June 15, 2026. The model shares the same 1-trillion-parameter mixture-of-experts architecture as Kimi K2.6 with 32 billion active parameters per token, but is dedic The post details improvements over K2.6: +21.8% on Kimi Code Bench v2 (62.0 vs 50.9), +11% on Program Bench (53.6 vs 48.3), +31.5% on MLS Bench Lite (35.1 vs 26.7), 10% gains on agent execution benchmarks including MCP Atlas and MCP Mark Verified, and 30% fewer reasoning tokens. Examples shown include agentic coding, r

Winzheng reports that Moonshot AI announced the open-source release of Kimi-K2.7-Code, a coding-focused model that achieves roughly a 30% reduction in token usage while maintaining higher intelligent performance compared to prior versions. The article frames the release as a breakthrough in inference efficiency that sp The model is described as specifically designed for code generation and understanding, with improved multilingual support (especially Chinese code scenarios) and gains in accuracy on complex algorithm tasks alongside reduced computational resource consumption. Moonshot's open-sourcing move — making weights freely avail

Moonshot AI released Kimi K2.7-Code as a coding-focused, agentic model, shipping open weights on Hugging Face under a Modified MIT license and accessible via the Kimi API and Kimi Code. Moonshot reports a +21.8% gain on its own Kimi Code Bench v2 versus K2.6, and targets long-horizon software engineering with planning, tool use, and debugging. The MarkTechPost article specifies a 1T-parameter Mixture-of-Experts architecture that activates 32B per token, with 384 experts (8 routed plus 1 shared) across 61 layers, MLA attention, SwiGLU feed-forward, and a 400M-parameter MoonViT vision encoder. Native INT4 quantization is supported, context is 256K tokens, and thinking mode is mandatory. Sampling is fixed at temperature 1.0 and top-p 0.95, with a 32,768-token default output.