DeepSeek-OCR

Context compression, not just text recognition, is the research question behind DeepSeek-OCR: the model explores how far a vision encoder can compress a document page into a small number of visual tokens that a language model can still decode into faithful markdown. Resolution tiers span Tiny at 512x512 (64 vision tokens) through Large at 1280x1280 (400 tokens), plus a dynamic Gundam mode that mixes tile sizes for dense pages. Inference is built for scale: running under vLLM, the repository reports roughly 2,500 tokens per second on an A100-40G when batching PDF pages. The reference setup targets Python 3.12 with PyTorch 2.6 and CUDA 11.8 or newer, and an NVIDIA GPU is required. Code and weights are MIT licensed, so commercial use is unrestricted, and the repository collected about 23,700 stars after its October 2025 release; a DeepSeek-OCR2 follow-up landed in January 2026 in its own repository. Teams building LLM data pipelines use it to convert large document archives into training and RAG corpora cheaply.