Oh yeah. DeepSeek-OCR Model - Contexts Optical Compression. Best model in town for OCR.

DeepSeek Contexts Optical Compression.

The podcast provides the technical overview of the DeepSeek's Contexts Optical Compression technique, which addresses the challenge of processing long texts by treating pages as optical context and converting text into images. This methodology is implemented in the DeepSeek-OCR model, which utilizes a DeepEncoder and a DeepSeek-3B-MoE decoder to efficiently compress textual information into a small number of vision tokens; for example, roughly 1,000 text tokens are reduced to about 100 vision tokens. The underlying philosophy is that vision tokens serve as a compressed format for text-heavy pages, allowing the large language model (LLM) to decode from a small latent state rather than raw text sequences, significantly reducing the token budget and maintaining model speed. This system also enables an innovative feature called Dynamic Gundam mode which adds tiles and a global view to manage context, and it allows for progressive downscaling of older text history, functioning as an optical analogue of human memory fading. The results demonstrate that DeepSeek-OCR achieves state-of-the-art quality in optical character recognition (OCR) and document parsing with a fraction of the token budget compared to competing models.