Liquid AI has released the LFM2.5-VL-DSpark draft model, a speculative decoding drafter designed to accelerate inference for the LFM2.5-VL-3B vision-language model. The drafter captures hidden states from fixed layers of the target model to generate candidate tokens, adding only 280M parameters (8.9% overhead) while maintaining identical dimensionality across modalities.

  • Decoding speedups reach up to 3.13x on Apple M5 Max devices and 2.66x on NVIDIA H100 GPUs.
  • End-to-end latency improvements range from 1.56x to 2.62x on-device and 1.64x to 2.27x on GPU.
  • The model provides day-one integration support for llama.cpp, MLX-VLM, and SGLang.
  • Speculative decoding accelerates only the decode phase; prefill and vision encoding remain compute-bound bottlenecks on edge devices.

The release enables faster deployment of open-weight vision-language models across diverse hardware environments without restricting fine-tuning or distribution.