Liquid AI releases DSpark draft models, speeding up decoding by up to 4x

JosephJacks_ · x · 2026-08-21

Liquid AI has released DSpark draft models for the LFM2.5 series, utilizing speculative decoding to significantly accelerate inference speeds.

Mechanism: A lightweight draft model proposes candidate tokens, and the target model verifies them in a single forward pass, trading minimal memory increase for substantial speedups.
Performance Data:
On H100: LFM2.5-8B-A1B on MATH500 increased from 428 to 1362 tok/s (3.18x).
On-device (M4 Max): LFM2.5-1.2B-Instruct on HumanEval increased from 136 to 389 tok/s (2.87x).
LFM2.5-2.6B achieved 2.67x speedup on H100.
Use Cases: Ideal for on-device function calling applications or cloud workloads with critical latency requirements.

More from Models