DeepSeek OCR Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection. DeepSeek OCR is a vision language model for Optical Character Recognition with "Contexts Optical Compression." It supports native and dynamic resolutions, multiple prompt modes (document to markdown, free OCR, figure parsing, grounding), and can be run via vLLM (high throughput) or HuggingFace Transformers. It processes images and PDFs, outputting structured text or markdown. Installation Prerequisites CUDA 11.8+, PyTorch 2.6.0 Python 3.12.9 (via conda recommended) Setup Alternative: upstream vLLM (nightly) Model Download Model is available on HuggingFace: deepseek ai/DeepSeek OCR Inference: vLLM (Recommended for Production) Single Image — Streaming Batch Images PDF Processing (via vLLM scripts) Benchmark Evaluation Inference: HuggingFace Transformers Transformers Script Prompt Reference Use Case Prompt Document → Markdown <image \n< grounding Convert the document to markdown. General OCR <image \n< grounding OCR this image. Free OCR (no layout) <image \nFree OCR. Parse figure/chart <image \nParse the figure. General description <image \nDescribe this image in detail. Grounded REC <image \nLocate <\ ref\ TARGET TEXT<\ /ref\ in the image. Supported Resolutions Mode Resolution Vision Tokens Tiny 512×512 64 Small 640×640 100 Base 1024×1024 256 Large 1280×1280 400 Gundam (dynamic) n×640×640 + 1×1024×1024 variable Configuration (vLLM) Edit DeepSeek OCR master/DeepSeek OCR vllm/config.py : Common Patterns Process a Directory of Images Convert PDF Pages to Markdown Grounded Text Location (REC) Troubleshooting transformers version conflict with vLLM vLLM 0.8.5 requires transformers =4.51.1 — if running both in the same env, this error is safe to ignore per the project docs. Flash Attention build errors CUDA out of memory Use smaller resolution: base size=512 or base size=640 Disable crop mode=False to avoid multi crop dynamic resolution Reduce batch size in vLLM inputs Model output is garbled / repetitive Ensure NGramPerReqLogitsProcessor is passed to LLM — this is required for proper decoding: Tables not rendering correctly Add table token IDs to the whitelist: Multi GPU inference Key Files

deepseek-ocr