LiquidAI has released an experimental DSpark draft model for its LFM2.5-VL-3B vision-language model, adding a speculative-decoding path intended to accelerate inference without altering output quality. The drafter contributes approximately 280 million parameters—an 8.9 % increase over the 3 B target—and is described by the company as a simplified four-layer, attention-only model with a recommended block size of 8 or 9.
According to LiquidAI's own measurements on the MMSpec benchmark across six vision-based tasks, the drafter achieves decode speedups of up to 3.13× on an Apple M5 Max (via MLX) and up to 2.66× on an NVIDIA H100, with corresponding end-to-end latency improvements of up to 2.62× and 2.27× respectively. On a MacBook M3 Ultra running llama.cpp, decode speedups range from 1.57× to 2.14× and end-to-end gains from 1.30× to 1.77×. LiquidAI notes that because speculative decoding accelerates only the decode stage and not vision encoding or prefill, end-to-end gains are bounded by Amdahl's law, particularly on edge devices where prefill dominates wall time.
The model ships with day-one integration support for llama.cpp, MLX-VLM, and SGLang, with specific pull requests referenced for each runtime. LiquidAI states that speculative decoding is exact: the target model verifies every proposed token, so greedy output is identical to running the target alone.
The LFM2.5-VL-DSpark draft model is available on Hugging Face in both Safetensors and GGUF formats and is designated open-weight, with the company stating it can be downloaded, fine-tuned, and deployed without restrictions. The release is explicitly labelled experimental.
All performance figures cited above are vendor-reported measurements by LiquidAI; no third-