Kimi in practice: long-running agents with K3 and K2 on Runpod

Best practices from Moonshot AI for splitting work across Kimi models, cutting token use and running agents for hours, with a demo you can follow on Runpod.

Getting good results from an open model on real work depends on how you run it: which model leads, which ones do the legwork, and how you keep a long task from burning tokens or drifting off course.

In this 30-minute session from Runpod's monthly model series, Feihu from Moonshot AI and Chris Vela from Runpod share how teams use Kimi in practice. Many Kimi users pair K3 as the lead agent with K2-series models as subagents, and both are available on Runpod Public Endpoints today.