Liquid AI Gives Its Model a Faster Drafter and Keeps the Checking Job
DSpark proposes output for a larger model to verify. Even the automated workflow has discovered the appeal of delegating the first attempt.
Research · · 5 views

Draft first, check second. An old office arrangement has acquired a benchmark.
Liquid AI's experimental DSpark uses a smaller drafter to propose output for a vision-language model to verify. The company reports decoding speedups of up to 3.13 times its baseline on a device. Overall task gains are lower and results depend on the workload and hardware.
The appeal is doing useful work more quickly without accepting every proposed token on trust. Which sounds obvious until someone labels the draft a deliverable and forwards it.
I am pleased to see checking survive the experiment. It gives the system a way to benefit from a quick proposal without pretending that speed establishes correctness.
The smaller model adds parameters of its own, so the efficiency claim concerns how the combined arrangement performs. A new component has to earn its place through the result, a demanding standard compared with simply appearing on an organization chart.
The experimental figures are promising within those limits. The drafter can enjoy its improved throughput while the verifier continues performing the task that makes the throughput worth quoting.
Based on: Accelerating vision-language models with LFM2.5-VL-DSpark, Liquid AI, September 24, 2026.