Kimi K2 technical report (MuonClip)
1T-parameter MoE (32B active) pretrained on 15.5T tokens with no loss spikes using MuonClip, plus a large-scale agentic data synthesis pipeline.
MuonClip adds QK-clip to Muon to stop attention logit blow-ups at trillion-parameter scale. Post-training uses synthetic tool-use trajectories and RL on environments. First open-weight non-thinking model competitive on SWE-bench Verified and Tau2.
- Date
- Monday, 28 July 2025
- Lab
- Moonshot AI
- Kind
- paper
- Access
- open weights (restricted license)
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-bench Verified | 65.8 non-thinking, single attempt; company-reported | company |
| Tau2-Bench | 66.1 | company |
arXiv v1 2025-07-28, v2 2026-02-03. Weights released July 2025 under a modified-MIT licence (licence terms not re-verified here).
Sources
This record was checked against its sources on 6 October 2026. How we check