Liquid AI has released a DSpark draft model for its LFM2.5-VL-3B vision-language model, designed to accelerate decoding throughput on GPUs and edge devices. The update provides day-one support for llama.cpp, SGLang, and MLX-VLM, claiming significant speed improvements for vision-based tasks.
This release addresses the high computational cost of vision-language models by accelerating the decode phase without sacrificing accuracy. For the streaming and AI ecosystem, this enables more responsive visual analysis on consumer-grade hardware like Apple's M5 Max and M3 Ultra chips. While prefill remains a bottleneck due to Amdahl's law, the significant gains in decoding throughput allow for higher concurrency and lower per-user latency in multi-turn visual conversations. Watch for whether these speculative decoding gains can be maintained as Liquid AI expands support to quantized models and larger batch sizes in production environments.
The integration with inference frameworks like llama.cpp and SGLang highlights a broader industry shift toward optimizing local execution for multimodal AI.
Liquid AI has released the DSpark drafter, a speculative decoding model for the LFM2.5-VL-3B vision-language model. This update significantly improves decoding throughput, achieving speedups of up to 3.13x on edge devices and 2.66x on NVIDIA H100 GPUs, enabling more responsive visual analysis on consumer-grade hardware without sacrificing output quality.
The DSpark drafter is a speculative decoding model designed by Liquid AI to accelerate the decode phase of the LFM2.5-VL-3B vision-language model.
The 280M parameter drafter model adds only 8.9% to the total parameter count of the LFM2.5-VL-3B model.
Day-one integration for the DSpark drafter is available for llama.cpp, SGLang, and MLX-VLM inference frameworks.
The model was trained exclusively on AMD hardware using Liquid AI's proprietary framework.
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source