Tag: Speculative Decoding

A small drafter model makes Liquid AI’s LFM2.5 run three times faster

Speculative-decoding checkpoints from Liquid AI boost LFM2.5 throughput by up to 3.18x without changing outputs.