Run Qwen2-VL on Your Computer with Text, Images, and Video, Step by Step

Your local multimodal chat model

Example of inference by Qwen2-VL 2B

Alibaba’s Qwen2-VL are vision language models now available as 2B and 7B parameter models. They are generative language models that support multimodal inputs. You can provide Qwen2-VL, along with text, a single image, multiple images, or a 20-minute video!

The models demonstrate impressive performance in visual understanding. Like Microsoft’s Florence-2, Qwen2-VL can perform many types of tasks such as OCR, image captioning, question answering, visual grounding, etc.

Florence-2: Run Multitask Vision-language Models on Your Computer

Quantized versions, using the AWQ and GPTQ formats, were also published by Alibaba to facilitate deployment on smaller GPUs.

In this article, I first review Qwen2-VL’s architecture and performance. Then, we will see how to use Qwen2-VL with multiple images and videos using small GPUs (8 GB and 12 GB). I explain how to set up the model and format the prompt step by step.

The examples detailed in this article are also implemented in this notebook: