antirez streams Kimi K3.1 on a 128GB machine via expert prefetching with residual predictions

antirez demonstrated streaming Kimi K3.1 on a 128GB machine by using the previous layer's residuals to run the next MoE router ahead of time, enabling expert prefetching that almost always predicts the correct experts.

2026-11-05 ~ 2026-11-05 · 2 related posts

• antirez: prefetching MoE experts via previous-layer router lets K3.1 stream on a 128GB box — antirez · 2026-11-05
• antirez prefetches MoE experts via router trick to run Kimi K3.1 on a 128GB machine — antirez · 2026-11-05