Accelerating vision-language models with LFM2.5-VL-DSpark

A new DSpark draft model significantly accelerates the LFM2.5-VL-3B vision-language model using an advanced technique called speculative decoding. This method uses a small, efficient "drafter" to predict multiple tokens at once, achieving decoding speedups of up to 3.13x on consumer devices without altering the final output. The performance boost comes at a minimal cost, increasing the model's parameter count by just under 9%. While decoding is much faster, the overall end-to-end speedup is limited by the initial image encoding and prompt processing stages, which are not accelerated by this method.