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

A vision-language model draft model has been released that speeds up inference by using speculative decoding, a technique that trades a small increase in memory for faster processing without changing output quality. The drafter adds approximately 280 million parameters, increasing the target model's size by 8.9%, while delivering decoding speedups up to 3.13x on device and up to 2.66x on GPU hardware. Speculative decoding works by having the draft model capture hidden states from the target model and propose multiple candidate tokens at once, which the target model then verifies. The speedup matters most for the decoding phase, though overall improvements are more modest when vision encoding and initial processing stages already consume significant time, a limitation explained by how different computational bottlenecks affect different parts of the inference pipeline.