Kimi K2.7 Code

Kimi K2.7 Code supports OpenAI-compatible API. Key usage patterns:\n\n1. Simple Chat: Standard chat.completions.create() with reasoning content in response.\n2. Chat with Image: Pass image\u005furl with base64-encoded images.\n3. Chat with Video: Pass video\u005furl with base64-encoded videos (experimental, official API only).\n4. Preserve Thinking: Full reasoning content is retained across multi-turn interactions.\n5. Interleaved Thinking + Multi-Step Tool Call: See K2 Thinking documentation.\n\nFor detailed code examples, see the HuggingFace model card.

1. Coding-focused agentic model built upon Kimi K2.6\n2. 30% reduction in thinking-token usage compared to K2.6\n3. Native INT4 quantization (same method as Kimi-K2-Thinking)\n4. Forced thinking + preserve\u005fthinking mode for enhanced multi-turn coding agent performance\n5. Multimodal: Supports text, image, and video input\n6. MLA attention mechanism with SwiGLU activation\n7. MoonViT vision encoder (400M params)\n8. Kimi Code CLI as recommended coding agent framework (https://www.kimi.com/code)\n9. Open source under Modified MIT License\n10. Available via OpenAI/Anthropic-compatible API on platform.moonshot.ai

Context Length: 256K (262,144 tokens)\nInput Modalities: Text, Image, Video\nOutput Modalities: Text\n\nNote: Chat with video content is an experimental feature and is only supported in the official API for now (not in third-party APIs deployed with vLLM or SGLang).

Kimi K2.7 Code forces thinking and preserve\u005fthinking as True. This feature is enabled by default and cannot be disabled.\n\nPreserve thinking retains full reasoning content across multi-turn interactions and enhances performance in coding agent scenarios.\n\nRecommended settings for Thinking mode:\n- Temperature: 1.0\n- Top-p: 0.95\n- Instant mode is not supported\n\nK2.7-Code also supports Interleaved Thinking and Multi-Step Tool Call (same design as K2 Thinking).

Kimi-K2.7-Code API is available on https://platform.moonshot.ai with OpenAI/Anthropic-compatible API.\n\nRecommended inference engines:\n- vLLM\n- SGLang\n- KTransformers\n\nKimi-K2.7-Code has the same architecture as Kimi-K2.5/Kimi-K2.6, and the deployment method can be directly reused. The version requirement for transformers is >=4.57.1, <5.0.0.\n\nDeployment examples: https://huggingface.co/moonshotai/Kimi-K2.7-Code/blob/main/docs/deploy_guidance.md

Architecture: Mixture-of-Experts (MoE) Total Parameters: 1T Activated Parameters: 32B Number of Layers (Dense layer included): 61 Number of Dense Layers: 1 Attention Hidden Dimension: 7168 MoE Hidden Dimension (per Expert): 2048 Number of Attention Heads: 64 Number of Experts: 384 Selected Experts per Token: 8 Number of Shared Experts: 1 Vocabulary Size: 160K Context Length: 256K Attention Mechanism: MLA Activation Function: SwiGLU Vision Encoder: MoonViT (400M parameters)

Kimi K2.7 Code is a coding-focused agentic model built upon Kimi K2.6. With substantial improvements on real-world long-horizon coding tasks, it strengthens end-to-end task completion across complex software engineering workflows while improving token efficiency, reducing thinking-token usage by approximately 30% compared with Kimi K2.6.