DSpark Draft Models Accelerate LFM2.5 Inference
Liquid AI has released DSpark draft checkpoints for three models in its LFM2.5 family: LFM2.5-1.2B-Instruct, LFM2.5-2.6B and LFM2.5-8B-A1B. The approach adds a speculative-decoding path that uses small draft models to propose tokens before the target model verifies them. Each draft model has about 300M parameters and requires only a modest memory increase.
In tests with SGLang on one H100 80 GB, average speedups were 2.67x for LFM2.5-2.6B, 2.10x for LFM2.5-1.2B-Instruct and 2.54x for LFM2.5-8B-A1B. The highest measured GPU result was 3.18x. With llama.cpp and Metal on an M4 Max MacBook Pro, average improvements were 2.27x, 2.54x and 1.18x respectively. Liquid AI also reports a 57% average reduction in function-calling latency for LFM2.5-2.6B.
The release includes support for llama.cpp and SGLang, with checkpoints available in Safetensors and GGUF formats. Under greedy decoding, rejected draft tokens are replaced by the target model’s output, preserving the baseline sequence and benchmark accuracy.
Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.