Qwen-VL

Qwen-VL is a large vision-language model from Alibaba Cloud. According to its Hugging Face model card, it accepts images, text and bounding boxes as input and outputs text and bounding boxes. The page covers the pretrained Qwen-VL and the chat-tuned Qwen-VL-Chat, including an Int4 variant. Its language backbone is Qwen-7B, and the associated paper is arXiv 2308.12966, published August 24, 2023.

Key features

• Grounding: referring expression comprehension; the card says it generalizes to Chinese grounding zero-shot without Chinese grounding training data.
• Text reading: text-oriented VQA across TextVQA, DocVQA, ChartQA, AI2D and OCR-VQA, helped by a 448 input resolution.
• Dialogue: multilingual conversation and multi-image interleaved chat in Qwen-VL-Chat.
• Vendor claims: state-of-the-art zero-shot captioning on Flickr30K and strong RefCOCO results among generalist models of similar scale; Qwen-VL-Chat scores 645.2 (English) and 401.2 (Chinese) on TouchStone, according to the card.
• License: free for research and development, with commercial use also allowed; the card links a license file and a commercial application form.

Use cases

• Image captioning and visual question answering.
• Locating objects from text descriptions with bounding boxes.
• Reading text in documents, charts and screenshots.

Pricing and availability

The weights are free to download from Hugging Face. The card says use is free for research and development and commercial use is allowed, with a license file and a form for commercial applications; read them before deployment. No hosted price is listed on the card. We checked it on 2026-10-29. The card does not state a parameter count outright; its benchmark tables label the model Qwen-VL-7B, consistent with the Qwen-7B backbone.

Getting started

1. Open the model card and install the dependencies it lists.
2. Load Qwen-VL for pretrained use or Qwen-VL-Chat for dialogue; an Int4 chat variant is available for lower memory.
3. Pass an image plus a text prompt, and request boxes when you need grounding.
4. Test against newer Qwen vision models on your own images before adopting.

FAQ

What does Qwen-VL accept as input?

Images, text and bounding boxes, and it outputs text and bounding boxes.

What is its language backbone?

Qwen-7B, according to the model card.

Can I use it commercially?

The card says it is free for research and development and commercial use is allowed, with a license file and a commercial application form.

What replaced Qwen-VL?

The card names no successor. For a later Qwen vision model in this catalog, see Qwen2-VL 72B Instruct.

Alternatives

• Qwen2-VL 72B Instruct: a later Qwen vision-language model in this catalog.
• LLaVA 1.6: an open vision-language model.
• CogVLM Base 490 HF: another open vision-language model.

Conclusion

Qwen-VL is an early, well-documented open vision-language model with grounding and text-reading skills. Read the model card, compare it with newer Qwen vision models, and browse more in the models category.