The Pulse
Liquid AI Adds DSpark Decoding to LFM2.5-VL-3B
Liquid AI’s team article describes an experimental draft model for faster decoding.
AI.info Team ·
Liquid AI released an experimental speculative-decoding model for its LFM2.5-VL-3B vision-language model on September 24, 2026. The add-on, called LFM2.5-VL-DSpark, is designed to make the model generate text faster while leaving the target model in charge of verifying proposed tokens. Liquid AI says the draft model adds about 280 million parameters, or 8.9% to the 3-billion-parameter target.
DSpark drafts; LFM2.5-VL-3B verifies
Speculative decoding uses a smaller drafter to propose several tokens at once. The larger target model checks those candidates in a single pass, accepting matches and supplying its own token when a candidate fails verification. That can reduce the number of costly steps in text generation without asking the smaller model to replace the target.
For the vision-language version, Liquid AI says the drafter captures hidden states from selected layers of LFM2.5-VL-3B. Image patches and text tokens are projected into a shared representation before those layers, so the drafter works with hidden-state vectors of the same dimensionality for either input modality. The inference algorithm remains the same as for the company’s text DSpark models.
The drafter has four attention-only layers. Liquid AI says it selected that design after ablations across three, four and five layers, and trained it for 10 epochs on a mixture of vision-language supervised fine-tuning data weighted toward expected workloads. The company reports that acceptance improved with additional training tokens before reaching diminishing returns. Its component breakdown lists 193 million parameters for the decoder stack, 21 million for hidden-state projection, 65.5 million for the Markov head and 6,400 for norms and the confidence head, for a total of 279.5 million.
The speed gains vary by device and task
Liquid AI tested the system on six vision-language workloads: general visual question answering, text-focused visual question answering, image captioning, chart questions, complex reasoning and multi-turn conversation. The evaluations used a DSpark block size of eight and followed the MMSpec benchmark.
On an M5 Max using MLX, decoding ran 2.30× to 3.13× faster across tasks, while end-to-end latency improved by 1.56× to 2.62×. On an M3 Ultra running llama.cpp, decoding gains ranged from 1.57× to 2.14×, with end-to-end improvements of 1.30× to 1.77×.
The H100 figures in the article conflict. Its summary says decoding speedups are up to 2.66×, while the detailed results give a range of “20.4x to 2.66x.” The article does not explain the difference, so the H100 decoding result cannot be stated consistently from those figures. For H100 end-to-end latency, the detailed results report improvements from 1.64× to 2.27×.
Image processing limits overall acceleration
DSpark speeds up decoding, not image encoding or prompt prefill. A vision-language model must first process an image through its vision encoder, then handle the visual tokens and text prompt. Liquid AI says those steps can account for a substantial share of total response time, particularly on edge devices with less computing capacity than datacenter GPUs. That helps explain why end-to-end improvement can be smaller than the decode speedup.
The draft model is available in Safetensors and GGUF formats, with support for llama.cpp, MLX-VLM and SGLang. Liquid AI notes that users need compatible builds for the integrations. The target verifies every proposed token; the article says this makes speculative decoding exact, with greedy output matching the target model running alone. The result is a shorter decode stage paired with a 279.5-million-parameter draft model, rather than a faster vision encoder.