Liquid AI ships a DSpark drafter for vision models

Liquid AI released an experimental DSpark draft model for its LFM2.5-VL-3B vision-language model. The drafter speculates on upcoming tokens from the target model’s hidden states and the target verifies them, which trades a small amount of memory for decoding speed without changing output. Liquid reports decoding throughput gains of up to 2.66x on GPUs and 3.13x on edge devices, with end-to-end gains of up to 2.27x and 2.62x.

What’s actually new

• Speculative decoding extended to vision-language models. Once tokens reach the hidden layers, text and image patches are just tensors, so the same drafting algorithm used for Liquid’s text drafter applies unchanged to a VLM.
• A small price for the speedup. The drafter is about 280M parameters, raising the deployed model’s parameter count by 8.9%, and uses four layers with a recommended block size of 8 or 9.
• Day-one runtime support. GGUF checkpoints for llama.cpp, optimized paths for Apple Silicon through MLX-VLM, and SGLang for GPU serving, with weights in Safetensors and GGUF on Hugging Face.
• Trained entirely on AMD hardware, using Liquid’s own training framework, then measured on their Pipette benchmark infrastructure across six MMSpec vision tasks.

Why it matters

Liquid is candid about the ceiling here, which is more useful than the headline number. Speculative decoding only accelerates the decode phase, and on edge hardware a VLM’s vision encoding and prefill dominate wall time, so a 3x decode win turns into a much smaller end-to-end gain by Amdahl’s law. Liquid’s own numbers show the gap: 2.30x to 3.13x faster decoding on an M5 Max but only 1.56x to 2.62x end-to-end, and 1.57x to 2.14x decoding on an M3 Ultra. The output is lossless, but higher sampling temperatures cut the acceptance rate and with it the throughput benefit.

(Original source: Liquid AI)