Morning Edition · Saturday, October 31, 2026Published at 2:16 AM EDT · New York

Ant Group's Inclusion AI Reports 5.8x to 6.9x Decoding Speedups From a Parallel Draft-Tree Method Called UBTree

The drafter builds trees of up to 64 candidate tokens that the target model checks in a single forward pass, and it beat the DARTree baseline in all 28 reported comparisons.

Inclusion AI, the open-model group at Ant Group, has released UBTree, a speculative decoding method that builds the draft tree in parallel rather than one token at a time (AI ML Big Data). Standard tree drafters extend each branch autoregre…