Kimi K3 gets open-sourced speculative decoding draft models

TorchSpec open-sourced the K3 Draft Collection—three SOTA speculative decoding draft models (EAGLE-3, DFlash2, DSpark) for Kimi K3—trained on NVIDIA GB200 in collaboration with vLLM.

2026-09-03 ~ 2026-09-04 · 2 related posts