Multimodal large language models (MLLMs) combine text with images or video. In conversations, all visual tokens must be processed before the first answer word appears, causing long waits.
Raincheck processes only a fraction of tokens upfront and defers the rest to later steps, using idle compute during generation and after the answer. It recovers lost information without slowing down the response.
In tests, Raincheck recovered 77.7% of the quality lost by permanently discarding tokens and reduced initial wait by up to 24.1%, with almost no increase in per-token time.