← Back to Discover

DeepSeek-VL2

Open family of vision-language models from DeepSeek that answers questions about images, reads documents, tables and charts, and locates described objects. Three sizes, with weights on Hugging Face.

Open models
Visit website ↗

What it does

DeepSeek-VL2 is a family of Mixture-of-Experts vision-language models. It targets tasks that mix images and text, such as visual question answering, optical character recognition, understanding documents, tables and charts, and visual grounding. There are three variants: DeepSeek-VL2-Tiny, DeepSeek-VL2-Small and DeepSeek-VL2, with 1.0B, 2.8B and 4.5B activated parameters respectively. Weights for all three are listed for download on Hugging Face, each with a sequence length of 4096. The GitHub repository adds installation steps, an inference script that handles single and multiple images, and a Gradio web demo. Its notes warn that the example script may need about 80GB of GPU memory with the Small model and more with the full one, and it describes an incremental prefilling option for 40GB cards. The code is MIT licensed, while use of the models is subject to the DeepSeek Model License, which the page says supports commercial use.

Capabilities

  • Mixture-of-Experts vision-language models
  • Visual question answering and OCR
  • Document, table and chart understanding
  • Visual grounding of objects named in text
  • Tiny, Small and full-size variants
  • Inference script and Gradio demo in the repository

When to use it

  • You want an open vision-language model to run on your own GPU
  • You need to pull information out of document, table or chart images
  • You want to locate an object in a picture from a text description
  • You are comparing multimodal models for research or a commercial project