Liquid AI's DSpark draft model speeds up LFM2.5-VL-3B inference by up to 3.13x on device

On September 24, Liquid AI released an experimental DSpark draft model for its vision-language model LFM2.5-VL-3B. Called LFM2.5-VL-DSpark, the draft model adds a speculative decoding path that trades a minimal memory increase for faster inference while keeping output quality unchanged. According to the company, decode speed improves by up to 3.13x on-device and 2.66x on an H100, with end-to-end gains up to 2.62x and 2.27x.

The drafter follows the same architecture as Liquid AI's text LFM2.5-DSpark drafters. It captures hidden states from fixed tapped layers of the target model and uses them to draft a block of k candidate tokens. Image patches and text tokens are projected into a shared representation before those layers, allowing the same inference algorithm to work across modalities. Based on ablations over 3, 4, and 5 layers, the team selected a 4-layer attention-only drafter with a block size of 9, trained for 10 epochs on a vision-language SFT mixture. The drafter adds about 280M parameters, an 8.9% increase over the 3B target.

Day-one integrations include llama.cpp, MLX-VLM, and SGLang, and the model is available in Safetensors and GGUF formats. Liquid AI notes that speculative decoding only accelerates the decode phase, not vision encoding or prefill; when those stages dominate latency, end-to-end gains are limited by Amdahl's law.