D
DeepSeek-OCR
Description
DeepSeek-OCR is an open model with about 3 billion parameters that reads text from images. It comes from a research paper on compressing text into images ("Contexts Optical Compression"). The example prompts convert a document page to Markdown or do plain text recognition. It runs with Hugging Face Transformers on NVIDIA GPUs; the authors tested Python 3.12 and CUDA 11.8 and list flash-attn among the dependencies, and an official vLLM support was announced shortly after the release. The weights are under the MIT license and are not gated. The page carries a multilingual tag but does not list languages.
Pros and cons
Pros
- MIT license, commercial use allowed
- Open weights without an access request
- Outputs Markdown from a document page
- Official vLLM support
Cons
- Needs an NVIDIA GPU
- The model page does not list supported languages
- Handwriting is not mentioned on the model page
Key facts
- License
- MIT
- Commercial use
- Allowed
- Hardware requirements
- NVIDIA GPU (the authors tested with CUDA 11.8)
- Developer
- DeepSeek
- Checked on
- Oct 20, 2026
Editorial score
How strongly the AI Expertise editors recommend this open model for each of its tasks right now, on a 0–10 scale. The score combines output quality, capabilities, ease of launch, license openness and freshness.