NVIDIA releases 4B speculative decoding draft boosting Qwen3.8-27B acceptance by 46%

_akhaliq · x · 2026-11-18

NVIDIA has published a speculative decoding draft model on Hugging Face: a 4B-parameter draft designed to accelerate Qwen3.8-27B, improving token acceptance by 46% over the model's built-in drafting head. It's an official option for speeding up inference in self-hosted deployments.

Original post →

More from Infra

Infra channel →