Kimi K3
2.8T-parameter open-weights MoE (104B active, 1M context) on Kimi Delta Attention and Attention Residuals, reported third on Artificial Analysis behind two closed models at launch.
Largest open model so far, with 896 experts (16 active), KDA hybrid attention, MXFP4 weights with quantization-aware training from SFT onward, and about 2.5x scaling efficiency over K2. Moonshot says it still trails Claude Fable 5 and GPT-5.6 Sol. API-first; weights followed 11 days later under a custom license.
- Date
- Thursday, 16 July 2026
- Lab
- Moonshot AI
- Kind
- open-weights
- Access
- open weights (restricted license)
- Price
- $3.00 input / $15.00 output per M tokens ($0.30 cache hit), API, 2026-07
Figures
| Measure | Value | Measured by |
|---|---|---|
| Total / active parameters | 2.8T / 104B 93 layers, 896 experts, 16 active | company |
| GPQA-Diamond | 93.5 | company |
| Terminal-Bench 2.1 | 88.3 Kimi Code harness | company |
| BrowseComp | 91.2 context compaction at 300K; 90.4 with 1M context and no management | company |
Announced 2026-07-16 on Kimi apps/API; weights released 2026-07-27 under the Kimi K3 License (separate agreement for Model-as-a-Service operators above $20M revenue over 12 months; Wikipedia reports revenue sharing up to 30%). Fortune says 2.7T; Moonshot says 2.8T. Benchmarks are company-run; the 'Artificial Analysis #3' ranking is from Wikipedia and unverified here.
Sources
- www.kimi.com/blog/kimi-k3
- huggingface.co/moonshotai/Kimi-K3
- simonwillison.net/2026/Jul/27/kimi-k3/
- fortune.com/2026/07/16/moonshots-kimi-k3-pushes-chinese-ai-into-fable-level-territory/
This record was checked against its sources on 6 October 2026. How we check