Kimi K2.7-Code: open-source coding model with better token efficiency
3 hours ago
- #AI Model
- #Deployment Guide
- #Coding Agent
- Instructions for using moonshotai/Kimi-K2.7-Code include libraries (Transformers), notebooks (Google Colab, Kaggle), local apps (vLLM, SGLang, Docker Model Runner).
- Kimi K2.7 Code is a coding-focused Mixture-of-Experts (MoE) model with 1T total parameters, 32B activated, 256K context length, and improved coding agentic performance.
- Evaluation shows improvements over Kimi K2.6 in benchmarks like Kimi Code Bench v2, Program Bench, MLS Bench Lite, and agentic tasks such as Kimi Claw 24/7 Bench.
- Deployment options include vLLM, SGLang, KTransformers; requires transformers version >=4.57.1, <5.0.0.
- Model usage examples cover chat completion with thinking mode, visual content (images/videos), preserve_thinking mode, interleaved thinking, multi-step tool calls, and integration with Kimi Code CLI.
- Released under Modified MIT License; contact support@moonshot.ai for questions.