DSpark makes a vision model up to 3.13 times faster, but cannot speed up the first token

Liquid AI has released an experimental DSpark drafter for the LFM2.5-VL-3B vision-language model. The smaller model proposes blocks of upcoming tokens, while the target model verifies them and accepts only valid proposals. With matched sampling settings, the resulting distribution remains equivalent to running the target model alone.

An 8.9 percent parameter overhead multiplies decoding speed

DSpark has 279.5 million parameters, four layers and a recommended block size of 8 or 9 tokens. Beside the 3-billion-parameter target, it increases deployed parameter count by 8.9 percent. Liquid AI provides support in llama.cpp, MLX-VLM and SGLang, with weights available in Safetensors and GGUF formats.

On an M5 Max, decoding improved by 2.30 to 3.13 times depending on the task, while end-to-end latency improved by 1.56 to 2.62 times. On one H100 80 GB, decoding improved by 2.04 to 2.66 times and the full run by 1.64 to 2.27 times.

Edge deployments gain speed without replacing the target model

For teams already serving LFM2.5-VL-3B, the attraction is the narrow scope of the change. The drafter does not replace the target or trade quality for quantization. It proposes candidates that the original model checks. An integration can therefore shorten image descriptions, document answers or visual conversations without changing application logic.

The benefit grows when a response contains enough output tokens. Faster decoding accumulates over a long caption or multi-turn exchange. For a short answer, work outside DSpark's reach may dominate the request.

Vision encoding and prefill still collect the latency bill

An image must first pass through a vision encoder, after which the target model processes hundreds of visual tokens. Speculative decoding accelerates neither this stage nor time to first token. That is why the best end-to-end gain is lower than the headline decoding figure. The published tests also use 16-bit weights, batch size 1 and six MMSpec tasks. Acceleration for quantized models is outside this release.

Short answers and concurrent traffic will provide the useful test

Real application profiles will decide the outcome: vision-encoding share, prompt length, response length and request concurrency. Liquid AI measured an advantage at higher concurrency too, although the gap narrowed as concurrency rose. Teams should measure end-to-end latency and time to first token rather than pinning 3.13 times alone to a dashboard.

Lilith's verdict

DSpark seats a small prompter beside three billion parameters and lets the main model run faster. The camera still waits at the starting line to process the image, so the stopwatch belongs on the whole race, not only the finishing straight.