Moonshot releases Kimi K3 at 2.8 trillion parameters
A multimodal Mixture-of-Experts model with 896 experts and a 1M context window.
Why it matters: It pushed the open-weight size ceiling into the trillions and set up the K3.1 line.
- Parameters
- 2.78T total
- Active per token
- 104B
- Weights size
- 1,561 GB (HF repo)
- Context
- 1,048,576 tokens
- License
- Custom: Kimi K3 license
| Benchmark | Score | Type | Measured or published by |
|---|---|---|---|
| GDPval-AA v2 | 1686 Elo | Cited | Moonshot, citing Artificial Analysis |
| DeepSWE | 67.5 | Self-reported | Moonshot |
| Terminal-Bench 2.1 | 88.3 | Self-reported | Moonshot |
Sources: huggingface.collmgateway.io