Kimi K2.7 Code

Moonshot AI announced Kimi K2.7 Code on June 16, 2026, positioning it as a coding-focused agentic model tuned for long-horizon, end-to-end programming tasks. The official forum post reports state-of-the-art benchmark results with tools, including Kimi Code Bench V2 at 62.0 (vs. K2.6 at 50.9), Program Bench at 53.6 (vs. The announcement details several model-level improvements: higher success rates on 4,000+ tool call sequences and 12+ hours of continuous execution, with generalization across Rust, Go, and Python; roughly 30% fewer reasoning tokens on average compared with K2.6 to reduce overthinking; native multimodal tool use includ

GitHub added Moonshot AI's Kimi K2.7 Code to Copilot's model picker on July 1, 2026, making it the first open-weight model offered as a selectable option inside Copilot, according to Tech Insider Ireland. The integration is available to Copilot Pro, Pro+, and Max subscribers, with support across more than nine client a The rollout happened shortly after Moonshot AI, a Beijing-based lab, published the Kimi K2.7 Code weights to Hugging Face on June 12, 2026, under a Modified MIT license, with the gap between open-weight release and Copilot general availability reported at 19 days. GitHub's official account reposted the announcement wit

Lorphic's July 13, 2026 explainer walks through the Kimi K2 model family, situating K2.7 Code as a code-specialized member of the same MoE architecture shared by K2, K2.5, and K2.6. It describes the foundation as a 1-trillion-parameter Mixture-of-Experts model with 32B active parameters per token, using 384 routed expe The article positions K2.7 Code among version-by-version differences, benchmark data, and API considerations, helping developers understand which variant they are working with and how the naming maps to distinct capabilities and licensing implications. While derivative of Moonshot's primary sources, the explainer conso

eesel AI's third-party review confirms that Kimi K2.7 Code is Moonshot AI's coding-specialized model built on Kimi K2.6, designed for long-horizon software engineering tasks like planning, multi-file editing, tool use, and multi-step debugging in a single session. The architecture is a 1-trillion-parameter Mixture-of-E The review highlights a notable design constraint: thinking mode cannot be disabled and every request runs the model's full chain-of-thought, with the API erroring if users try to override temperature, top-p, or penalty parameters away from fixed defaults. Moonshot frames this as improving multi-step tool-calling relia

GitHub announced on July 1, 2026 that Kimi K2.7 Code is generally available in GitHub Copilot, making it the first open-weight model offered as a selectable option in the Copilot model picker and giving developers a lower-cost coding option. The model is hosted by GitHub on Microsoft Azure and is billed at provider lis Specific client version requirements are stated for selection in the model picker: Visual Studio Code 1.127.0 or later, Visual Studio 17.14.6 or later, Copilot CLI, the GitHub Copilot cloud agent, the Copilot App, github.com, GitHub Mobile (iOS and Android), JetBrains 1.9.1-251 or later, Xcode, and Eclipse. For Copilot

Mehmet Özel's Medium analysis frames Kimi K2.7 Code not as a larger model but as a more disciplined one — still within the Kimi K2 family's roughly 1T total / 32B active MoE regime and aimed at long-context, tool-using, agentic workloads. The piece emphasizes that the meaningful change over K2.6 is reasoning-token disc The article positions K2.7 Code as a better-trained engineer rather than a new brain, targeting long-horizon coding workflows with controlled reasoning and efficient execution. It is an opinion/analysis piece whose core factual claims mirror Moonshot's own messaging — the same architectural family, the same focus on ag

AIToolsReview's June 21, 2026 review confirms Kimi K2.7 Code as Moonshot AI's open-weights coding model released on June 12, 2026. It reiterates the architecture: a 1 trillion parameter MoE with 32 billion active parameters, 384 experts with 8 active, a 256K context window, thinking-only inference mode, and Modified MI Coding benchmark scores attributed to Moonshot are Kimi Code Bench v2 at 62.0, Program Bench at 53.6, and MLS Bench Lite at 35.1 — described as sharply up from K2.6. One agentic benchmark result, MCP Mark Verified at 81.1, is cited as ahead of Claude Opus 4.8's 76.4 but behind GPT-5.5's 92.9. Pricing is listed at appro

CoreWeave has published a technical write-up confirming that Moonshot AI's Kimi K2.7 Code, the latest coding-focused agentic model in the Kimi K2 family, is now running on its serverless inference platform. According to the post, K2.7 Code is built on the same trillion-parameter Mixture-of-Experts architecture as K2.6, The article details CoreWeave's Blackwell-era optimization work for K2.7 Code, centered on NVFP4 quantization and a DFlash speculator, developed jointly by its Applied Training and Inference teams. It positions this as a follow-on to the K2.6 optimization that previously earned CoreWeave top placement on Artificial Ana

This deep dive documents Moonshot AI's release of Kimi K2.7-Code as open weights on June 12, 2026 via Hugging Face, the Kimi Open Platform API, Kimi Code CLI, and Cloudflare Workers AI on day one. The architecture is detailed as a 1T-total / 32B-active Mixture-of-Experts model with 384 routed experts plus one shared ex Moonshot self-reports +21.8% improvement over K2.6 on its own Kimi Code Bench v2 and a 30% reduction in reasoning tokens, though the page notes all public benchmarks are Moonshot's own proprietary suites and independent SWE-bench Verified / Pro / FrontierCode scores were not yet available as of June 15, 2026. The page

Moonshot AI released Kimi K2.7-Code as a coding-focused, agentic model whose weights ship on Hugging Face under a Modified MIT license, also reachable through the Kimi API and Kimi Code. The architecture is a 1-trillion-parameter Mixture-of-Experts model that activates 32B parameters per token, using 384 experts (8 sel On Moonshot's published benchmarks, K2.7-Code beats the prior K2.6 on every reported row; the largest coding jump is Kimi Code Bench v2, rising from 50.9 (K2.6) to 62.0 (K2.7-Code), with the release noting a +21.8% improvement. The comparison table also lists GPT-5.5 at 69.0 and Claude Opus 4.8 at 67.4 on the same benc