Kimi K2.7-Code Release Notes: Open-Source Coding Model with Enhanced Token Efficiency

Moonshot AI has released Kimi K2.7-Code, an open-weights coding-focused agentic model built upon the Kimi K2.6 architecture. The model is specifically designed to improve end-to-end task completion in complex software engineering workflows and significantly reduce operational costs by decreasing thinking-token usage by approximately 30% compared to its predecessor.

Model Architecture and Technical Specifications

Kimi K2.7-Code utilizes a Mixture-of-Experts (MoE) architecture to balance high capacity with inference efficiency. It features a total of 1 trillion parameters, with 32 billion parameters activated per token.

Key Technical Specs

Specification Detail
Architecture Mixture-of-Experts (MoE)
Total Parameters 1T
Activated Parameters 32B
Context Length 256K
Attention Mechanism MLA
Number of Experts 384 (8 selected per token)
Vocabulary Size 160K
Vision Encoder MoonViT (400M parameters)

Performance Benchmarks

Kimi K2.7-Code demonstrates substantial improvements over Kimi K2.6 across coding and agentic benchmarks, narrowing the gap with frontier models like GPT-5.5 and Claude Opus 4.8.

Coding Performance

Benchmark Kimi K2.6 Kimi 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

Agentic Performance

Benchmark MCP Atlas MCP Mark Verified Kimi Claw 24/7 Bench
Kimi K2.6 69.4 72.8 42.9
Kimi K2.7 Code 76.0 81.1 46.9
GPT-5.5 79.4 92.9 52.8
Claude Opus 4.8 81.3 76.4 50.4

Core Features and Capabilities

Thinking Mode and Token Efficiency

Kimi K2.7-Code forces a preserve_thinking mode. This ensures that the full reasoning chain is retained across multi-turn interactions, which is critical for maintaining context in complex coding agent scenarios. Despite this forced reasoning, the model is 30% more efficient in its thinking-token usage than K2.6.

Multimodal Input

The model supports image and video input via the MoonViT vision encoder, allowing it to process visual context alongside code.

Native INT4 Quantization

To facilitate deployment on consumer or limited hardware, Kimi K2.7-Code adopts a native INT4 quantization method, consistent with the Kimi-K2-Thinking model.

Deployment and Usage

Recommended Inference Engines

For optimal performance, Moonshot AI recommends the following engines:

  • vLLM
  • SGLang
  • KTransformers

API Configuration

The model provides an OpenAI/Anthropic-compatible API. Recommended settings for Thinking mode include a temperature of 1.0 and a top_p of 0.95. Note that "Instant mode" is not supported.

License

The model weights and code are released under a Modified MIT License. Community discussion indicates this license is similar to the original MIT license but includes a requirement to advertise/credit Moonshot AI if the model is used in a commercial product.

Community Insights and Real-World Application

Users on Hacker News have reported practical successes and specific technical observations:

  • Complex Patching: One user reported using Kimi K2.7-Code to successfully rebase a 177KB OpenSSL patch from version 3.3.1 to 3.5.7, noting the agent performed nontrivial work to handle the patch not applying cleanly.
  • Tool Call Adherence: Users noted that K2.7-Code follows custom tool call formats more reliably than K2.6, indicating improved instruction following.
  • Cost-Efficiency: Discussion highlighted the competitive pricing of Chinese models compared to US-based frontier models, though some users expressed concerns regarding data privacy and the geopolitical implications of using models from Chinese providers.

Sources