DeepSeek-OCR

Open weights, runs locally, no vendor API: DeepSeek-OCR keeps document parsing on your own hardware. DeepSeek reports matching or beating heavier OCR models on OmniDocBench. Less lock-in, more control over your data. deepseek-ai/DeepSeek-OCR

DeepSeek-OCR turns docs, tables and screenshots into text using about 100 visual tokens per page, per the paper. 3B MoE with ~570M active params, open weights, runs on a single GPU. Useful for local invoice and receipt parsing. deepseek-ai/DeepSeek-OCR

DeepSeek 开源了一个名为 DeepSeek-OCR 的 OCR 模型，核心创新是“上下文光学压缩”——把整页文档或图像压缩成极少量的视觉 token（模型可理解的“图像片段”），再交给大语言模型处理，在探索视觉文本压缩极限的同时大幅降低 token 开销。 deepseek-ai/DeepSeek-OCR t.co

DeepSeek-OCR: Contexts Optical Compression - 24k Stars, 2.2k Forks DeepSeek-OCR is a powerful tool for visual-text compression, with over twenty-four thousand GitHub stars and two thousand two hundred forks. It uses advanced machine learning to convert images into text, making it perfect for document processing and OCR tasks. For example, it can convert a PDF document into searchable text, saving you time and effort. To get started, visit the project page on GitHub or follow the installation instructions in the README. Explore the boundaries of visual-text compression and join the community for support and updates. Find more details and the latest version in the repository. Link in first comment 👇 #DeepSeekOCR #OCR #Compression 🔗 deepseek-ai/DeepSeek-OCR