Kimi K2.7 Code

Moonshot AI released Kimi K2.7-Code on June 12, 2026 as a coding-focused, agentic model with weights on Hugging Face under a Modified MIT license and access via the Kimi API and Kimi Code. The architecture is a 1T-parameter Mixture-of-Experts with 32B active per token, 384 experts (8 selected plus 1 shared), 61 layers, MLA attention, and SwiGLU feed-forward, plus a 400M-parameter MoonViT vision encoder for image and video input. Context is 256K tokens (262,144), with native INT4 quantization for serving, and thinking mode is mandatory — disabling it returns an API error. Sampling is constrained to temperature 1.0, top-p 0.95, n 1, penalties 0.0, with a default max output of 32,768 tokens, and the model can be self-hosted using vLLM for long-horizon software engineering workflows.

Kimi K2.7 Code is positioned by Moonshot AI as the company's most intelligent coding model to date, with stronger instruction following on long-context tasks and a higher success rate on software engineering work. The page states the official API model ID is kimi-k2.7-code, with a 256K context window and support for text, image, and video inputs across chat and agent task types. Main entry points listed are Kimi Code, the Kimi API, Claude Code, Cline, and RooCode, with tool support including tool calls, JSON mode, partial mode, and automatic context caching. Thinking mode is always enabled, and disabling it is not supported, reinforcing the model's coding-and-agent orientation rather than general chat use.

GitHub announced on July 7, 2026 that Kimi K2.7 Code, an open-weight model, is now generally available in GitHub Copilot and is the first open-weight option in the Copilot model picker. After the July 1, 2026 announcement for Pro, Pro+, and Max plans, availability has now been extended to Copilot Business and Copilot Enterprise plans with usage-based billing at provider list pricing. Kimi K2.7 Code is hosted by GitHub on Microsoft Azure and is off by default for Copilot Business and Enterprise, so plan administrators must enable the Kimi K2.7 Code policy before any organization member can select it. GitHub recommends that administrators review open-weight models against their own security, compliance, and data-governance requirements before enabling.

Kimi K2.7 Code is described as a 1-trillion-parameter open-weight coding agent with 32B active per token and a 256K context window, supporting native image and video input. According to the page, it reports +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, and +31.5% on MLS Bench Lite versus K2.6, while using 30% fewer reasoning tokens for long-horizon software engineering and MCP tool use. The model is positioned for multi-file refactors across 10+ languages with repo-scale 256K context and forced, preserved thinking that cannot be disabled. The page lists four capability pillars — coding, forced thinking, MCP/agentic tool use, and multimodal input — and notes that a 6× high-speed serving mode is announced by Moonshot for throughput on long agentic runs.

BenchLM's Kimi Code Bench v2 leaderboard, verified September 10, 2026, lists Kimi K2.7 Code by Moonshot AI at 62.0% overall score (65.49 secondary metric) with a 256K context window, explicitly naming the model as open weights and ranking it behind Kimi K3's 72.9%. The page describes Kimi Code Bench v2 as a Moonshot AI in-house benchmark for realistic coding-agent tasks across backend, infrastructure, systems, security, frontend, and ML/data engineering, with a 2026 quarterly refresh cadence, providing dated, model-specific benchmark evidence for Kimi K2.7 Code.

Kimi's official "What's New" changelog documents the September 2026 rollout of the K2.8 Preview model release across Kimi Code products, explicitly noting that it delivers coding and agent capability improvements with significantly more efficient thinking than Kimi K2.7 Code while keeping the same Model ID (kimi-for-co The same changelog describes Kimi Code CLI v0.43.0 (Sep 14, 2026) and v0.42.0 (Sep 9, 2026) features including automatic AI session titles, Ctrl-X session deletion, fairer Goal Mode time budgets, generally available Remote Control, a sortable media preview bar, and HEIC/HEIF/BMP image support, providing dated Kimi Code