LiquidAI extends lossless speculative decoding to vision-language models

JosephJacks_ · x · 2026-09-25

LiquidAI shows how their speculative-decoding approach for text LFMs extends naturally to vision-language models.

Key insight: by the time information reaches the language backbone's hidden layers, text tokens and image patches are both tensors, so the original modality no longer matters to the drafter. A lightweight drafter uses LFM2.5-VL-3B's hidden states from different layers to speculate several tokens ahead, verified by the target model in a single forward pass.

Under matched sampling settings the decoding is distribution-equivalent to direct sampling, making the output lossless.

Original post →

More from Models

Models channel →