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