Daniel Walsh demonstrates an open-weight vision-language model derived from DeepSeek-OCR on a long PDF. He contrasts conventional OCR, structured document parsers and vision-language models, emphasizing that each serves different document needs.
Daniel Walsh explains how optical compression reduces the input token count and how reference sliding window attention retains the source image while restricting attention to recent generated text. The approach aims to prevent the growing output cache from slowing long transcriptions.
Daniel Walsh notes the model’s 32,000-token input limit, differences between supported resolution modes and the limitations of its benchmark comparison. He argues that parallel page-by-page processing may often be preferable, while a single pass can help when content spans page boundaries.
Watch on YouTube
