Kimi-K2-Base
Kimi K2 Base is the pretrained foundation of the original K2 release. Its mixture of experts contains one trillion parameters but activates about 32 billion for each token.
MuonClip controls unstable attention updates
K2 was pretrained on 15.5 trillion tokens. Its MuonClip optimizer extends Muon with QK clipping to control training instability. The report describes a stable training run and releases both the base and post-trained checkpoints.Kimi K2: Open Agentic Intelligence
Sparse computation still uses a large checkpoint
The specification lists 384 experts, eight selected experts per token and one shared expert. It uses Multi-Head Latent Attention across 61 layers. The active-parameter figure therefore describes token computation, not the total set of weights.Kimi-K2-Base model card
The added condition concerns attribution
Commercial products above 100 million monthly users or US$20 million monthly revenue must display Kimi K2 under the license's additional condition. The threshold is monthly, and the condition is not the later K3 model-service license.Kimi K2 Base and Instruct license
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | July 2025 |
| Access | Downloadable |
| Model identifiers | moonshotai/Kimi-K2-Base |
| Context window | 131,072 tokens Published checkpoint limit; this is not a promise of continued API access. Kimi-K2-Base checkpoint configuration |
| Parameters | 1 trillion total; 32 billion active per token.Kimi-K2-Base model card |
| Training stage | Pretrained base checkpoint for adaptation.Kimi-K2-Base model card |
| Weight license | Modified MIT, with a large-commercial-product attribution condition.Kimi K2 Base and Instruct license |
Related models
- Kimi-K2.5
Moonshot AI · Kimi K2
Multimodal K2 successor that combines image and video understanding with coding and reasoning.
Launch date Downloadable - Kimi-K2-Instruct
Moonshot AI · Kimi K2
Original instruction-tuned K2 checkpoint for tool use and coding without extended thinking mode.
Launch date Downloadable