FOLLOW A THREAD

推测解码

Connect scattered updates into a story you can understand.

ProjectsNews10/29, 18:46

Ant Group open-sources the UBTree speculative decoding solution, with average test speedup increased to 6.88 times

Ant Group’s Bailing open-sourced the UBTree speculative decoding solution, improving large-model inference efficiency through parallel draft-tree selection. In the Qwen3-4B seven tests, the average speedup increased from 5.95 times to 6.88 times.

ProjectsNews10/21, 19:18

Eagle releases EAGLE-4: Qwen inference accelerated by 3.6 times, switches to a restrictive commercial license

Eagle releases EAGLE-4 speculative decoding solution; reports indicate that on 4 NVIDIA B300 GPU, it raises Qwen3.8-27B's single-request generation speed to 863.8 Token/second. The new version has released its code publicly, but its commercial license is more restrictive than those of the previous generations.

# Technology Developments # EAGLE-4 # Speculative decoding # Large-model inference # Performance testing

All loaded2 shown · 12 per batch