Moonshot AI Kimi K2.7 Code Overview

Model Architecture and Performance

Kimi K2.7 Code is a Mixture-of-Experts model with 1 trillion total parameters and 32 billion activated parameters per token. It is specifically optimized for long-horizon software engineering tasks, such as multi-file refactoring and repository-wide debugging. The model utilizes a mandatory thinking mode that reportedly reduces thinking tokens by approximately 30 percent compared to previous iterations, leading to faster response times. Moonshot AI reports a score of 62 on their internal Kimi Codebench v2, an improvement over the 50.9 score achieved by the K2.6 model.

Agentic Workflow and CLI

The Kimi Code CLI provides a terminal-based interface for repository exploration and file modification. Key operational features include:

• Plan Mode: Activated via Shift + Tab, this mode forces the agent to analyze the codebase and propose a structured approach before executing any file changes.
• Session Management: Users can manage context with commands like /compact to compress conversation history, /sessions to resume previous work, and /model to switch between model versions.
• Read-Only Default: The agent defaults to read-only operations for repository exploration, requiring explicit confirmation before executing shell commands or modifying files.
• Compatibility: The membership API key is compatible with third-party coding agents including Claude Code, Roo Code, and OpenCode.

Pricing and Quotas

Moonshot AI offers subscription plans that provide a weekly refreshed quota, intended to offer more predictable costs for heavy agentic users compared to standard token-based API billing. The API itself remains available for occasional use at a rate of $0.19 per million cached input tokens, $0.95 per million uncached input tokens, and $4.00 per million output tokens. Subscription tiers range from $19 to $199 per month, with higher tiers providing increased concurrency caps and request limits.