LiquidAI has released an experimental DSpark draft model for its LFM2.5-VL-3B vision-language model, achieving substantial performance gains through speculative decoding. The new model delivers decoding speedups of up to 3.13x on Apple silicon and 2.66x on NVIDIA H100 GPUs, with end-to-end latency improvements reaching 2.62x and 2.27x respectively. The drafter adds only 280M parameters—representing an 8.9% memory increase—while preserving identical output quality to the base model.
The DSpark approach works by capturing hidden states from the target model at specific decoder layers and conditioning a simplified four-layer attention-only drafter to generate token candidates in blocks of 8-9. Speculative decoding then verifies each proposed token, ensuring greedy output matches the original model. LiquidAI has secured day-one support across three major inference frameworks: llama.cpp, MLX-VLM, and SGLang, making the models immediately usable by practitioners.
Released as open-weight models on Hugging Face in Safetensors and GGUF formats, the variants enable unrestricted downloading, fine-tuning, and deployment. While decode speedups are significant, LiquidAI notes that on edge devices, vision encoding and prefill processing consume most execution time, limiting overall end-to-end gains relative to GPU scenarios where decode is the primary bottleneck—a constraint imposed by Amdahl's law.
Key Points
Achieves up to 3.13x decoding speedup on Apple silicon and 2.66x on H100, with 2.62x and 2.27x end-to-end latency improvements respectively