Advanced coding model by MoonshotAI
kimi-k2.7-code
256K tokens

Kimi K2.7 Code:Built for Coding Agents

Kimi K2.7 Code is MoonshotAI's most capable coding model to date, tuned for long-horizon software engineering, tool calling, and agent workflows.

Built on the trillion-parameter K2 MoE backbone with 32B active parameters, Kimi K2.7 Code supports a 256K context window, text/image/video input, automatic context caching, and always-on thinking for multi-step engineering work.

Kimi K2.7 Code Signals
Kimi Code Bench v2
Official model-card score
62.0
Program Bench
Binary-to-program benchmark
53.6
MLS Bench Lite
ML systems benchmark
35.1
Context Window
Long-context coding
256K

What is Kimi K2.7 Code?

Kimi K2.7 Code is MoonshotAI's coding-focused K2 model for code generation, codebase understanding, tool use, and agentic software engineering. It is positioned as Kimi's strongest coding model to date.

Long-Horizon Coding
Designed to complete end-to-end programming tasks reliably over long contexts and multi-step workflows.
Always-On Thinking
Thinking mode is enabled by default and cannot be disabled, prioritizing planning and verification for engineering tasks.
Multimodal Engineering Input
Accepts text, images, and video so screenshots, UI recordings, logs, and requirements can feed coding tasks.
ToolCalls and JSON Mode
Supports ToolCalls, JSON Mode, Partial Mode, and automatic context caching for agent and API workflows.
K2.7 Code Technical Specs
  • • 256K context window for codebases and logs
  • • Always-on thinking; disabled thinking is not supported
  • • Text, image, and video input
  • • ToolCalls, JSON Mode, Partial Mode, context caching
  • • 1T / 32B active MoE architecture, modified MIT weights
K2.7 Code Use Cases
  • • Long-context codebase understanding
  • • Debugging, refactoring, and test generation
  • • Screenshot-to-UI and video-based issue analysis
  • • Claude Code, Cline, RooCode, and Kimi Code workflows
  • • Custom AI coding assistants via API

About Kimi K2.7 Code

Kimi K2.7 Code is a coding-specialized model built on the Kimi K2.6 foundation. It targets real software engineering workflows: codebase understanding, long-context instruction following, multimodal requirements, tool calls, and agentic task execution.

The release keeps the K2 family MoE backbone: 1T total parameters, 32B active parameters, 384 experts, 61 layers, MLA attention, SwiGLU activation, and a 256K token context window. The Hugging Face model card also lists MoonViT as the vision encoder for image and video workflows.

For developers, the practical entry points are Kimi Code, the Kimi API, OpenRouter, and compatible coding tools. K2.7 Code is most relevant when the task involves debugging, refactoring, test generation, multi-file edits, project analysis, or long-running coding agents.

Kimi K2.7 Code FAQ

Answers to common questions about Kimi K2.7 Code, Kimi Code, context length, thinking mode, pricing, and access.

What is Kimi K2.7 Code?

Kimi K2.7 Code is MoonshotAI's coding-focused K2 model for long-context software engineering, agent workflows, tool calling, and multimodal coding tasks. The official API docs describe it as Kimi's most capable coding model to date.

What is the Kimi K2.7 Code model ID?

The API model ID is kimi-k2.7-code. On Hugging Face, the model is listed as moonshotai/Kimi-K2.7-Code. On OpenRouter, the model ID is moonshotai/kimi-k2.7-code.

How long is the Kimi K2.7 Code context window?

Kimi K2.7 Code supports a 256K token context window, giving coding agents room for project files, logs, test output, and multi-turn engineering plans.

Can Kimi K2.7 Code disable thinking mode?

No. The Kimi API documentation states that Kimi K2.7 Code does not support non-thinking mode, and requests that disable thinking will result in an error.

Does Kimi K2.7 Code support images and video?

Yes. Kimi K2.7 Code supports text, image, and video input. This is useful for UI screenshots, design references, error screenshots, and interaction recordings.

What improved from Kimi K2.6 to Kimi K2.7 Code?

The Hugging Face model card reports higher scores for K2.7 Code than K2.6 on Kimi Code Bench v2, Program Bench, MLS Bench Lite, Kimi Claw 24/7 Bench, MCP Atlas, and MCP Mark Verified. It also reports about 30% lower thinking-token usage than K2.6.

Is Kimi K2.7 Code open weight?

Yes. The Hugging Face model card lists Kimi K2.7 Code under the modified MIT license, with model weights available as moonshotai/Kimi-K2.7-Code.

Where can I use Kimi K2.7 Code?

You can use it through Kimi API, Kimi Code, Hugging Face, OpenRouter, and compatible coding tools such as Claude Code-style Anthropic endpoints, Cline, and RooCode.