LFM2.5-VL-DSpark adds speculative decoding for faster vision inference
Liquid AI has released an experimental DSpark drafter for its LFM2.5-VL-3B vision-language model. The addition uses speculative decoding: a smaller model proposes blocks of tokens, which the target model verifies. According to Liquid AI, this preserves output quality while adding 280 million parameters, or 8.9% to the 3B model’s size.
The drafter has four attention-only layers and a recommended block size of eight or nine, depending on hardware. Tests across six vision tasks, including visual question answering, captioning, chart analysis, reasoning and multi-turn dialogue, followed the MMSpec benchmark. On an M5 Max using MLX, decoding was 2.30 to 3.13 times faster, while end-to-end latency improved by 1.56 to 2.62 times. On an H100, decoding reached up to 2.66 times the baseline speed, with end-to-end gains of up to 2.27 times.
Liquid AI said the model is supported from launch by llama.cpp, MLX-VLM and SGLang. The approach accelerates token generation but not image encoding or prompt processing, so those stages limit total gains, especially when they dominate latency. The model is available in Safetensors and GGUF formats.
Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.