ArchitectureTrainingReinforcement LearningAgents
Kimi K2: Open Agentic Intelligence
Moonshot AI·July 28, 2025
Kimi Team
View on arXivTL;DR
A 1.04T-parameter MoE (32B active) trained on 15.5T tokens for agentic and tool-use tasks, introducing the MuonClip optimizer for stable trillion-scale training plus a large agentic-data synthesis and joint-RL pipeline.
Why it matters
A landmark open-weight agentic model; MuonClip drew major attention as a route to stable training at trillion-parameter scale.