Tag

edge-inference.

2 writings found

Latest Archives

Speculative Decoding Brings VLMs to the Edge

Liquid AI's DSpark drafter achieves 3.13x speedups on-device by applying speculative decoding to vision-language models, with day-one support for llama.cpp and MLX.

Speculative Decoding Meets Vision: LFM2.5-VL Gets the Speed Treatment

Liquid AI's DSpark draft model accelerates vision-language inference by 2.3x on-device and 20x on GPU, but Amdahl's law reveals the real bottleneck.
View all rollups →