DeepSeek has unveiled a new open-source breakthrough with the release of DeepSeek-OCR, a novel model that tackles the growing challenge of processing long documents in large language models by transforming text into images for efficient compression. This innovation aims to significantly reduce computational costs and memory usage when handling lengthy textual content—issues that have become major bottlenecks in real-world AI applications. Traditional large language models struggle with long documents because their performance degrades rapidly as input length increases. Processing thousands or even tens of thousands of text tokens demands substantial compute power and memory, making such tasks impractical for many use cases. DeepSeek-OCR introduces a fresh approach inspired by human visual perception: instead of feeding raw text directly into the model, it first renders the document as a high-resolution image. Then, a visual encoder compresses the image into a much smaller set of visual tokens, which are subsequently decoded into text by a language model. This method isn’t just another OCR tool—it functions as a visual preprocessor designed to compress thousands of text tokens into just a few hundred visual tokens, effectively turning the document into a compact, machine-readable format. The model is built around two core components: DeepEncoder, a visual encoder, and DeepSeek-3B-MoE-A570M, a 3-billion-parameter mixture-of-experts (MoE) decoder with 5.7 billion activated parameters. DeepEncoder is engineered to handle high-resolution document images efficiently while minimizing memory usage and maximizing compression ratio. It combines two powerful vision architectures: SAM (Segment Anything Model), which excels at local, window-based attention and processes fine-grained visual details, and CLIP, which uses global attention to capture overall layout and semantic context. These two are connected via a 16x downsampled convolutional compression module. The process works in stages: SAM first extracts detailed features from the image, then the data is compressed before being passed to the computationally intensive global attention layers—reducing memory spikes and token explosion. On the decoding side, the MoE model reconstructs the original text from the compressed visual representation. To evaluate performance, DeepSeek tested DeepSeek-OCR on benchmarks like Fox and OmniDocBench. On English documents containing 600 to 1,300 text tokens, the model achieved accurate OCR with only 64 or 100 visual tokens—achieving compression ratios of up to 20x. At compression ratios below 10x, accuracy remained above 97%, and even at 20x, performance stayed around 60%. In OmniDocBench, DeepSeek-OCR outperformed other leading models such as GOT-OCR2.0 (256 tokens per page) and MinerU2.0 (over 6,000 tokens per page), delivering state-of-the-art results with far fewer visual tokens. Beyond standard text recognition, the model demonstrates strong "deep parsing" capabilities, thanks to its training on diverse document types including charts, chemical formulas, and geometric diagrams. It can convert graphs into tables, translate molecular structures into SMILES notation, and analyze geometric relationships—opening doors for applications in finance, scientific research, and education. DeepSeek has released the full code and model weights under an open-source license. According to technical reports, a single A100-40G GPU can process over 200,000 pages daily in production environments. Despite its promise, DeepSeek-OCR has limitations. Performance drops noticeably when compression ratios exceed 10x, likely due to information loss from over-compression or reduced image resolution. While it handles complex layouts well, extremely intricate page designs still pose challenges. Moreover, OCR evaluation focuses on perception and decoding accuracy, whereas real-world tasks like multi-turn dialogue involve deeper reasoning, memory retrieval, and contextual continuity. The model’s ability to preserve critical information across long sequences—especially in "needle-in-a-haystack" scenarios—remains untested. DeepSeek acknowledges these gaps and plans future work, including pre-training with mixed digital and optical text sequences and evaluating long-context retrieval accuracy. Still, DeepSeek-OCR marks a significant step forward—not just as a powerful OCR tool, but as a proof-of-concept for a new paradigm: using vision as a medium for compressing and reconstructing language. This approach could one day enable efficient handling of long conversation histories or massive knowledge bases by rendering them into compact visual representations. By bridging vision and language in this way, DeepSeek is paving the way for more scalable and intelligent AI systems.