FasterVLM introduces a training-free method to speed up Vision Language Models by pruning visual tokens using [CLS] attention, maintaining 90% performance while reducing computation by 95%.