Qwen2-VL 72B Instruct

Qwen2-VL-72B-Instruct is the instruction-tuned 72B model in Alibaba's Qwen2-VL series of vision-language models. Its Hugging Face model card lists it at 73B parameters in BF16, alongside smaller 2B and 8B variants. The card notes that a newer version, Qwen2.5-VL-72B-Instruct, is available. The associated paper, arXiv 2409.12191, was published September 18, 2024.

Key Features

• Dynamic resolution: images of arbitrary resolution map to a variable number of visual tokens. Users can set minimum and maximum pixel counts, rounded to multiples of 28, and the card's quickstart comment gives a default range of 4 to 16,384 visual tokens per image.
• Video: understands videos over 20 minutes long. It does not understand audio in videos.
• Languages: English and Chinese, plus text inside images in most European languages, Japanese, Korean, Arabic, Vietnamese, and others.
• Architecture: Multimodal Rotary Position Embedding (M-ROPE), which captures 1D text, 2D visual, and 3D video positions.
• Agent use: positioned for operating mobile phones and robots from visual input and text instructions.
• Vendor benchmarks: according to the model card, Qwen2-VL-72B scores 96.5 on DocVQA test, 70.5 on MathVista testmini, 86.5 on MMBench-EN test, and 73.6 on MVBench. The card compares against prior open-source results, Claude-3.5 Sonnet, and GPT-4o. These are vendor results.

License and Requirements

The license is listed as tongyi-qianwen. The card says to install the latest Hugging Face Transformers from source, otherwise you may see KeyError: 'qwen2_vl', and recommends the qwen-vl-utils package for handling visual inputs. The card does not state a context length.

Use Cases

• Document and chart question answering, where the card reports strong DocVQA results.
• Reading text inside images across many languages.
• Short and long video understanding without audio.

Limitations listed by the card: no audio support, image data updated only to June 2023, limited recognition of specific individuals and IP, weak complex instruction following, imprecise counting in complex scenes, and weak 3D spatial reasoning.

Getting Started

1. Install Transformers from source and qwen-vl-utils.
2. Load Qwen/Qwen2-VL-72B-Instruct, and set pixel limits to balance quality and memory.
3. Test on your own documents and video clips, and compare against Qwen2.5-VL if you can use the newer model.

FAQ

What is Qwen2-VL-72B-Instruct?

A 72B instruction-tuned vision-language model from Alibaba's Qwen team.

Can it understand video?

Yes, videos over 20 minutes, without audio.

Is there a newer version?

Yes. The model card points to Qwen2.5-VL-72B-Instruct. See Qwen-VL for an older Qwen vision-language entry in this catalog.

What is the license?

tongyi-qianwen, as listed on the card.

Alternatives

• Qwen-VL: the earlier Qwen vision-language model.
• Llama 3.2 Vision: Meta's vision-capable open model.
• Qwen3 VL Reranker: a newer Qwen vision-language reranking model.

Conclusion

Qwen2-VL-72B-Instruct is a capable open vision-language model, now superseded by Qwen2.5-VL per its own card. Read the model card for usage and see the models category for more.