LiquidAI's drafter decodes its vision model up to 3.13x faster
Victoria Shinder1 min readLiquidAI has released an experimental DSpark draft model for its vision-language model LFM2.5-VL-3B, published on Hugging Face on 24 September 2026. It follows the same pattern as the team's earlier LFM2.5-DSpark drafters: a speculative decoding path bolted onto an existing model, paying a minimal increase in memory footprint for a larger gain in speed, with output quality unchanged.
The figures given are decode speedups of up to 3.13x on device and 2.66x on an H100, and end-to-end gains of up to 2.62x and 2.27x respectively.
What it means
The gap between the decode number and the end-to-end number is the honest part of this release. 3.13x on the decode loop becomes 2.62x once everything else in the request is counted, because speculative decoding accelerates one stage and leaves image preprocessing, prefill and transport where they were. Anyone sizing a deployment should plan against the second number.
Speculative decoding is also a rare optimisation in that it does not ask you to accept a worse answer. Quantisation, distillation and pruning all trade quality for cost somewhere; a drafter either has its guesses accepted or it does not, and the verifier decides. That property is what makes "experimental" a reasonable label to ship under: the failure mode is a smaller speedup, not a different model.