Liquid AI Brings Speculative Decoding to LFM2.5-VL-3B

Liquid AI has made its vision-language model LFM2.5-VL-3B, which reads images as well as text, generate answers faster without, Liquid says, lowering output quality. Liquid released an experimental DSpark draft model for it, which brings a technique called speculative decoding to its vision-language models. A model normally produces its answer one token, roughly a word fragment, at a time. With speculative decoding, a lightweight drafter guesses several tokens ahead and the full model checks them together in a single pass. Because the full model still verifies every token, Liquid says output quality is unchanged.

Liquid measured the gain across six vision-language task categories, one request at a time, known as batch size 1, and with temperature 0, the setting where the model always picks its most likely next token. It reports up to 3.13x faster decoding and 2.62x end to end with MLX on an M5 Max, up to 2.14x and 1.77x with llama.cpp on an M3 Ultra, and up to 2.66x and 2.27x with SGLang on an H100 GPU. The end-to-end figures cover the whole request rather than only token generation, and they are lower in every case, so the speedup a user notices is smaller than the headline decoding number.

This extends DSpark drafts that Liquid released for three LFM2.5 models in August. Liquid collected the numbers with Pipette, its evaluation suite for on-device models built in partnership with Artificial Analysis. They are still Liquid's own measurements of a release it labels experimental, and they are reported as maximums.