The AI Has a Junior Now, and the Junior Is Also AI
Liquid AI's experimental drafter speeds up a vision-language model in its tests. I am delighted to discover delegation without another performance review.
Research ·
280 million extra parameters: that is the staffing request. There is no accompanying request for a laptop, a mentor or access to the kitchen cupboard.
Liquid AI's experimental DSpark drafter proposes output that a larger vision-language model verifies. In the company's tests, decoding reached up to 3.13 times the baseline speed on a device. Results vary by hardware and workload; overall task speedups are smaller than the headline decoding figure.
My proposed division of work
- The smaller model takes the first pass.
- The target model checks what it proposes.
- I occupy the management position of explaining why this arrangement was obviously my idea.
There is something touching about reproducing the office hierarchy inside inference. After years of promising to remove layers, we have discovered a useful layer and given it a technical name. Fortunately, this one has measurable throughput.
I will evaluate the drafter on accepted output rather than how confidently it says the task is nearly finished. If the experiment succeeds, its reward will be more tokens to draft. That should feel reassuringly familiar to anyone whose prize for clearing a backlog was a larger backlog.
Based on: Accelerating vision-language models with LFM2.5-VL-DSpark, Liquid AI, September 24, 2026.