DeepSeek-OCR

DeepSeek-OCR is the first generation of DeepSeek's document reader, published on Hugging Face on 17 October 2025 and last revised on 4 November 2025. The paper that explains it, "DeepSeek-OCR: Contexts Optical Compression" (arXiv 2510.18234, 21 October 2025), frames the model as an experiment rather than a product: if a page of text can be encoded as a small number of visual tokens and read back accurately, then an image is a cheaper container for long context than the text itself. The model has about 3.34 billion parameters — a DeepEncoder vision tower in front of a twelve-layer decoder in the DeepSeek-V2 mixture-of-experts style, with a language context of 8,192 tokens. What sets it apart from later document models is the explicit ladder of resolution modes, spelled out in the model card: Tiny at 512 pixels, Small at 640, Base at 1,024, Large at 1,280, and a "Gundam" mode that crops a 1,024-pixel page into 640-pixel tiles. Each rung trades visual tokens for fidelity, which is precisely the compression curve the paper measures. Two prompts drive it: free OCR, and a grounded conversion of the document to Markdown that keeps the layout. The weights are MIT-licensed and the model has no hosted API — DeepSeek never sold it, and the only way to use it is to run it, which it does in bfloat16 on a single graphics card. It also never stopped being used. Read on 25 August 2026, ten months after publication and seven months after its successor DeepSeek-OCR-2 appeared, the first generation still records 2.37 million downloads in thirty days — nearly twice the newer model, and more than any other model in DeepSeek's catalogue except DeepSeek-R1 and the V4-Flash line.

▸Videos

No videos yet.

▸Reviews