Liquid AI posts draft model, claims up to 3.13x faster decoding
The 280M drafter for LFM2.5-VL-3B adds 8.9% to the model's size, and Liquid AI says greedy output is unchanged.
Liquid AI published LFM2.5-VL-3B-DSpark on Hugging Face on Thursday, an experimental draft model that speeds up its 3-billion-parameter vision-language model through speculative decoding. The drafter is a 4-layer attention-only decoder of 280 million parameters, which the company puts at 8.9% on top of the 3B target.
On an M5 Max running MLX, Liquid AI reported decoding speedups of 2.30x to 3.13x and end-to-end gains of 1.56x to 2.62x. On an M3 Ultra with llama.cpp it reported 1.57x to 2.14x decoding and 1.30x to 1.77x end-to-end. On an H100 the reported decoding range is wider, 2.66x to 20.4x, with end-to-end gains of 1.64x to 2.27x.
All of those numbers are the company's own, measured on hardware and tasks it chose. Liquid AI said it tested across six vision tasks following the MMSpec benchmark, among them general and text VQA, image captioning, chart VQA and multi-turn conversation. No outside party has reproduced the figures.
DSpark works by tapping the target model's hidden states at a fixed set of layers and conditioning on them to draft a block of eight or nine candidate tokens. Liquid AI said the approach trades a minimal increase in memory footprint for a larger speedup without changing output quality.
The quality claim rests on how speculative decoding works rather than on a benchmark. Speculative decoding is exact, Liquid AI wrote, because the target model verifies every proposed token, so greedy output equals the target alone. Weights are on Hugging Face in Safetensors and GGUF, alongside the base LFM2.5-VL-3B.