LFM2.5-VL-DSpark makes vision-language inference up to 3.13x faster
Liquid AI’s LFM2.5-VL-3B can now decode up to 3.13x faster on an M5 Max, while end-to-end latency improves by as much as 2.62x. A 280M-parameter vision drafter predicts several next tokens from the target model’s hidden states, then the target checks them, adding only 8.9% to the deployed model and supporting llama.cpp, MLX-VLM, and SGLang.