Kimi K2.7 Code: Open-Source Agentic Coding Model
Kimi K2.7 Code is a coding-focused agentic model with improved long-horizon coding, stronger agent capabilities, and 30% lower thinking-token usage than K2.6.
TL;DR
- Kimi K2.7 Code is an open-source, coding-focused agentic model from Moonshot AI.
- It shows substantial improvements in coding benchmarks (+21.8% on Kimi Code Bench v2) and agentic task execution (+10% on Kimi Claw 24/7 Bench) compared to K2.6.
- The model reduces thinking-token usage by approximately 30% over K2.6.
- Kimi K2.7 Code is optimized for long-horizon coding tasks, improving instruction following and end-to-end success rates.
- It features a Mixture-of-Experts (MoE) architecture with 1 trillion total parameters and a 256K context length.
- The model weights are open-sourced on Hugging Face and it is available via Kimi Code and Kimi API.
- Kimi K2.7 Code requires thinking mode to be enabled and is purpose-built for coding, with K2.6 recommended for general-purpose tasks.
- Pricing for Kimi Code plans and Kimi API usage is detailed, with per-token billing for the API.