Liquid AI has released an experimental DSpark draft model for LFM2.5-VL-3B that accelerates response generation without changing the target model’s accepted output. The approach uses speculative decoding: a smaller model proposes blocks of tokens, and the full vision-language model verifies them before they are emitted.
The drafter has about 280 million parameters, adding 8.9% to the 3-billion-parameter target. In Liquid AI’s six-task tests, decoding ran 2.30 to 3.13 times faster on an M5 Max and 1.57 to 2.14 times faster on an M3 Ultra. End-to-end gains ranged from 1.56 to 2.62 times and 1.30 to 1.77 times, respectively. On an Nvidia H100, reported end-to-end gains ranged from 1.64 to 2.27 times.
The distinction between decoding and total latency matters. Images must first pass through a vision encoder, and the model must process visual and text tokens before producing its first answer token. DSpark does not accelerate those stages, so a large decoding gain may translate into a smaller overall improvement.
The open-weight drafter is available in Safetensors and GGUF formats, with integrations for llama.cpp, MLX-VLM and SGLang. Developers need compatible builds and must attach the draft model to LFM2.5-VL-3B when launching inference.