Qwen2-VL Models: 2B, 7B, 72B, Setup & Limits

Qwen2-VL is the Qwen Team’s 2024 vision-language family for text, image, and video understanding with text output. Its official 2B, 7B, and 72B Base and Instruct checkpoints remain publicly downloadable. Qwen2-VL is now a historical compatibility option rather than the latest Qwen vision family; for a current model decision, start with the Qwen Models hub.

Status checked August 3, 2026: the 72B Base and Instruct repositories are public; they are not API-only models. Downloadable checkpoint status is separate from any provider’s hosted API catalog, region, pricing, or retirement policy.

Qwen2-VL quick facts

Inputs: text, one or more images, and video frames or clips through a compatible processor.
Output: text. Visual grounding or coordinates are represented in generated text; the model does not generate an edited image or video.
No audio understanding: the official model card says audio inside a video is not understood.
Architecture changes: dynamic-resolution visual tokenization and Multimodal Rotary Position Embedding, or M-ROPE, for text, spatial, and temporal positions.
Declared model context: 32,768 positions in the official Qwen2-VL checkpoint configurations. Image and video tokens, text tokens, and generated tokens all consume practical capacity.
Family labels: 2B, 7B, and 72B. Hub interfaces may display slightly different total parameter counts, so use the exact repository ID when documenting or deploying a model.

Official checkpoints, roles, context, and licenses

Base checkpoints are pretrained models and are not the default choice for chat. Instruct checkpoints are post-trained to follow multimodal instructions. License terms are checkpoint-specific, especially at 72B.

Official checkpoint	Role	Declared context	License and status
`Qwen/Qwen2-VL-2B`	Base	32,768	Apache 2.0; public weights
`Qwen/Qwen2-VL-2B-Instruct`	Instruction-tuned	32,768	Apache 2.0; public weights
`Qwen/Qwen2-VL-7B`	Base	32,768	Apache 2.0; public weights
`Qwen/Qwen2-VL-7B-Instruct`	Instruction-tuned	32,768	Apache 2.0; public weights
`Qwen/Qwen2-VL-72B`	Base	32,768	Qwen License; public weights, non-Apache terms
`Qwen/Qwen2-VL-72B-Instruct`	Instruction-tuned	32,768	Tongyi Qianwen License; public weights, non-Apache terms

The two 72B repositories use named Qwen/Tongyi Qianwen license agreements rather than Apache 2.0. Read the exact license file in the checkpoint you plan to use; a family name is not a substitute for legal review.

What Qwen2-VL can do

Describe and answer questions about photographs, screenshots, charts, and documents.
Read text in images across the languages listed by the Qwen Team, including English, Chinese, Arabic, Japanese, Korean, Vietnamese, and multiple European languages.
Work with multiple images in one conversation when the processor and runtime preserve their order.
Sample and reason over video frames. The Qwen Team reported support for videos longer than 20 minutes, but that is a release claim, not an independent test result from this site.
Return textual descriptions, answers, or coordinate-like grounding data. Always validate structured output before using it in an automated workflow.

Current official Transformers example

The following compact image example follows the current official Qwen2-VL-7B-Instruct model page. Install a current compatible Transformers release and review its environment requirements before loading large weights.

from transformers import pipeline

pipe = pipeline(
    "image-text-to-text",
    model="Qwen/Qwen2-VL-7B-Instruct",
)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image",
                "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG",
            },
            {"type": "text", "text": "What animal is on the candy?"},
        ],
    }
]

result = pipe(text=messages)
print(result)

This snippet is documentation-derived and was not executed for this update. It intentionally contains no claimed output.

Serve Qwen2-VL with vLLM

The current official model page documents a standard vLLM server command:

pip install vllm
vllm serve "Qwen/Qwen2-VL-7B-Instruct"

Current stable vLLM documentation implements Qwen2-VL image and video processing. Older blanket claims that vLLM cannot handle Qwen2-VL video are obsolete. Actual video decoding, frame sampling, maximum duration, memory use, and OpenAI-compatible request format still depend on the installed vLLM version and server configuration.

Context and video limits in practice

The official configurations declare 32,768 positions, but that number is not a promise of 32,768 words or 32,768 text-only prompt tokens. Dynamic-resolution images can consume from a small number to many thousands of visual tokens, and video frame sampling can consume substantially more. Output tokens share the available sequence budget.

The “20 minutes+” statement comes from the Qwen2-VL release materials. It does not mean every 20-minute file fits unchanged in every runtime. Frame rate, resolution, processor limits, token budget, GPU memory, and requested output length all affect whether a specific clip can be processed.

Known limitations

Audio in video is not understood by Qwen2-VL.
The official card says its image dataset was current through June 2023, so newer visual entities or events may be missing.
Recognition of specific people, brands, and intellectual property can be incomplete.
Complex multi-step instructions, dense-scene counting, and three-dimensional spatial reasoning can fail.
OCR, coordinates, JSON, and other structured answers must be checked against the source image before automation.
Quantized builds and third-party conversions can change accuracy, supported modalities, maximum context, and runtime behavior.

Should you still use Qwen2-VL?

Use Qwen2-VL when you need to reproduce a 2024 result, maintain an existing deployment, or support an integration tied to an exact Qwen2-VL checkpoint. For a new vision project, compare later families first: Qwen2.5-VL, Qwen3-VL, and the current options summarized in the Models hub.

Use the Qwen download guide for official repositories and local-install choices. Hosted services are a different access lane: verify the exact provider ID and endpoint in the Qwen API guide, then check current costs and regional differences in the Qwen pricing guide.

Verification method and test limits

For this update, we checked the official Qwen2-VL collection, the six official Base and Instruct repositories, their configuration and license files, the Qwen Team’s release post, and current stable vLLM support. We did not download the weights, run GPU inference, call a hosted Qwen2-VL endpoint, or reproduce the Qwen Team’s benchmarks. Therefore, this page makes no original quality, latency, memory, or long-video result claim.

Frequently asked questions

Is Qwen2-VL-72B publicly downloadable?

Yes. Both Qwen/Qwen2-VL-72B and Qwen/Qwen2-VL-72B-Instruct are public official repositories. They use non-Apache Qwen/Tongyi Qianwen license terms that must be reviewed separately.

Does Qwen2-VL support video through vLLM?

Current stable vLLM code documents Qwen2-VL image and video modalities and a Qwen2-VL video backend. Pin and test the exact runtime version because decoding and sampling behavior can change.

Is Qwen2-VL the latest Qwen vision model?

No. It is a 2024 family retained here for history and compatibility. Use the Models hub to identify current hosted and open-weight choices.

Official sources

This is an independent technical reference and is not operated by Alibaba Cloud or the Qwen Team.

Qwen2-VL: Models, Licenses, Context, and Local Setup