Kimi K3
Open Weights
Moonshot AI
Moonshot AI's open-weight frontier model and the first open release in the 3-trillion-parameter class, at 2.8T total parameters with 104B activated. Built on Kimi Delta Attention and Attention Residuals with a Stable LatentMoE framework activating 16 of 896 experts, it has native vision and a 1M-token context window. Quantization-aware training from the SFT stage onward yields MXFP4 weights, cutting storage to roughly 1.4 TB against the ~5.6 TB FP16 would require.
Strengths
- First open-weight model in the 3-trillion-parameter class, at 2.8T total parameters
- Only 104B parameters activated per token via a Stable LatentMoE framework (16 of 896 experts)
- 1,048,576-token context window with native vision
- 93.5 on GPQA Diamond and 88.3 on Terminal-Bench 2.1
- Quantization-aware training from SFT onward gives MXFP4 weights at roughly 1.4 TB
- Roughly 2.5x the scaling efficiency of Kimi K2
Caveats
- Serving 2.8T parameters is a major infrastructure commitment even at MXFP4
- Released under the bespoke Kimi K3 License, not a standard OSI licence
- No release date is stated on the model card or repository
- Vision covers image understanding; the model card does not claim video
- Vendor
- Moonshot AI
- Context Window
- 1,048,576 tokens
- License
- Kimi K3 License
- Modalities
- textimage
| Benchmark | Score | Max |
|---|---|---|
| GPQA Diamond | 93.5 | 100.0 |
| Terminal-Bench 2.1 | 88.3 | 100.0 |
| BrowseComp | 91.2 | 100.0 |
| OSWorld-Verified | 84.8 | 100.0 |
| DeepSWE | 67.5 | 100.0 |
| MMMU-Pro | 81.6 | 100.0 |
Capabilities
Vision
Audio
Video
Tool Use
Resources
Pricing
Open weights under the Kimi K3 License (free to self-host); Moonshot's own hosted API publishes kimi-k3 at $0.30 per 1M input tokens (cache hit), $3.00 per 1M input tokens (cache miss) and $15.00 per 1M output tokens.