Guide
Choose the Kimi model and product surface while keeping consumer, API, coding-agent, personal-agent and browser permissions separate.
Published · Updated
Use K3 for broad frontier and native-vision work, K2.7 Code for coding-focused agent runs, Kimi API for application integration, Kimi Code for repository work, and Claw or WebBridge only after their broader account and browser permissions are justified.
The guide uses Moonshot's K3 launch, Kimi API pricing, Kimi Code repository and K2.7 Code model sources checked August 25, 2026.
K3 is Moonshot's broad 2.8T native-vision flagship with a 1M-token context. K2.7 Code is the coding-focused multimodal model with standard and HighSpeed routes. Keep required thinking or reasoning history intact in multi-step tool workflows.
Long-horizon coding, knowledge work, visual reasoning and broad frontier tasks.
Coding-specific agent workloads and high-speed serving options.
Kimi.com and Kimi Work expose user-facing creation workflows. Kimi API charges model input and output tokens, including extracted document content passed to a model. Consumer plan access does not imply API balance or rate limits.
Assistant, research, documents, slides, sites and sheets.
Metered developer access with model-specific prices and cache behavior.
Kimi Code is the repository and terminal coding runtime. Kimi Claw is a broader hosted personal-agent surface. Their execution, persistence and account permissions differ, so evaluate them with different security checklists.
WebBridge connects local agents to authenticated browser workflows. Use a dedicated browser profile, least-privilege sessions and explicit approval for consequential actions. Do not expose sensitive accounts merely because the model and runtime are trusted.
Audit cookies, account sessions, uploads and action approvals.
Keep Slides, Websites, Sheets and Docs within their user-facing product terms.