Kimi K3 open weights tested
Moonshot AI’s 1.56-terabyte open-weight model, without developer data about its Welsh ability.
The facts
| Developer | Moonshot AI |
|---|---|
| Released | 14 July 2026 |
| Architecture | MoE with 2.8 trillion total parameters / 104 billion active; Kimi Delta Attention and Attention Residuals; native visual capability; a one-million-token context window |
| Weights | 1.56 TB on Hugging Face |
| Licence | ‘Kimi K3 License’ — a bespoke permissive licence with commercial conditions; not a standard OSI licence |
Welsh: no claims or data
Chinese and English are the languages emphasised in K3’s documentation, and we found no claim, benchmark or reference to Welsh or another Celtic language. Its weights are openly available, unlike closed models, allowing local tests without relying on the developer’s API.
Its size is the practical barrier. With 104 billion active parameters and 1.56TB of weights, it exceeds most individual machines’ capacity. AIsteddfod’s realistic options are evaluation through an API service, or waiting for compressed versions.
CymraegBench v0.1 result
Overall score 75.12 · Welsh reasoning 77.92 · practical Welsh 72.33 · translation BLEU 67.62. Kimi placed ninth, the third-highest open-weight model after Muse Spark 1.2 and Qwen3.8 2.4T-A95B. See the full leaderboard.