DeepSeek OCR (github.com)

DeepSeek-OCR is an open-source OCR model suite released to probe how different vision encoders perform from an LLM-centric perspective. It’s targeted at researchers and engineers who want image-to-text/document understanding tightly coupled with large language models: supported workflows include direct image OCR, document→Markdown conversion, figure parsing, region referencing, and more via simple grounding prompts. The project integrates with both vLLM for high-throughput streaming and the Hugging Face transformers API (model keyed as deepseek-ai/DeepSeek-OCR), making it easy to test LLM+vision pipelines at scale. Technically, DeepSeek-OCR is designed for CUDA 11.8 + PyTorch 2.6.0 and uses performance primitives like flash attention and bfloat16 inference. It exposes multiple native resolutions with explicit vision-token budgets (Tiny 512² → 64 tokens, Small 640² → 100, Base 1024² → 256, Large 1280² → 400) and a “Gundam” dynamic-resolution mode (n×640² + 1×1024²) to trade off coverage vs detail. Example throughput: PDF processing ≈2,500 tokens/s on an A100-40G under vLLM. Installation requires a vLLM wheel, flash-attn, and careful dependency alignment with transformers. Overall, DeepSeek-OCR provides a pragmatic platform to evaluate resolution/tokenization strategies, dynamic cropping, and end-to-end LLM grounding for document understanding experiments.