Moonshot Kimi — Gated MLA MoE
Moonshot AI's Kimi family (Moonlight, Kimi-VL) uses a Gated MLA MoE architecture. When a major model drops, 1bit doesn't wait for vendor support — Kimi's architecture was reverse-engineered from the open weights and a 1BP converter was built.
Models
| Model | Params | 1BP Size | Backend(s) | Status |
|---|---|---|---|---|
| Kimi (Moonlight) | 16B (3B active) | — | GPU HIP | 🔄 arch analyzed, converter built |
| Kimi-VL | — | — | GPU HIP (vision) | 🔄 arch analyzed |
Notes
- Architecture reverse-engineered (Gated MLA MoE); converter built. Integration in progress.
- Same pure-C++ reverse-engineering process used for the NPU stack — no Python involved. See the reverse-engineering notes.
See also: engineering journey · full model support detail · all families