Liquid AI has released an experimental DSpark draft model for its LFM2.5-VL-3B vision-language model. The drafter uses speculative decoding: it proposes candidate tokens that the target model then verifies, so output quality is unchanged while inference is faster. The company says the drafter adds 280M parameters, an 8.9% increase over the 3B target.
Benchmarks on six vision tasks show decoding speedups of 2.30x to 3.13x with MLX on an M5 Max and 1.57x to 2.14x with llama.cpp on an M3 Ultra. On an H100, decoding improves by up to 2.66x. End-to-end gains are smaller: 1.56x to 2.62x on the M5 Max and 1.64x to 2.27x on the H100.
The post notes that speculative decoding only accelerates the decode phase. Vision encoding and prefill still take wall time, especially on edge devices, so the overall speedup is limited by Amdahl's law. The draft model uses four attention-only layers and a block size of 8 or 9 depending on hardware.
The drafter is available in Safetensors and GGUF formats with day-one support for llama.cpp, MLX-VLM, and SGLang. Liquid AI positions it as part of an open-weight family designed to run across edge and datacenter hardware.