2.8T · Stable LatentMoE 16/896 · KDA + AttnRes · vision/video · 1M ctx.
API kimi-k3. Weights ETA 2026-07-27. Active params undisclosed.
| piece | claim |
|---|---|
| MoE | 16 of 896 · Quantile Balancing |
| length | KDA (hybrid linear) |
| depth | AttnRes |
| vs K2 | ~2.5× scaling efficiency |
| QAT | MXFP4 W / MXFP8 A from SFT |
| thinking | always-on · effort=max at launch |
text|vision|video → KDA (length) + AttnRes (depth) → LatentMoE 16/896 → thinking → text/tools
Footgun: preserved thinking history required multi-turn, or quality collapses.
| K2.6 | K3 | |
|---|---|---|
| total | 1T | 2.8T |
| active | 32B | undisclosed |
| experts | 8/384+1 | 16/896 |
| ctx | 256K | 1M |
| attn | MLA | KDA+AttnRes |