Qwen-VL (2023): Models, Setup, Limits & Migration

Qwen-VL is the original 2023 vision-language model family released by the Qwen Team. This page covers two specific public checkpoints: the pretrained Qwen/Qwen-VL model and the aligned Qwen/Qwen-VL-Chat assistant. It is a historical and compatibility guide, not a page for the latest Qwen vision model or current hosted Qwen-VL product names.

Status checked August 3, 2026: both original repositories remain publicly downloadable. They use legacy custom Transformers code and a non-Apache Tongyi Qianwen license. New projects should compare later models through the Qwen Models hub.

What “Qwen-VL” means on this page

The Qwen Team released the original Qwen-VL and Qwen-VL-Chat checkpoints on August 22, 2023. The underlying language component was initialized from Qwen-7B, while the visual encoder was initialized from OpenCLIP ViT-bigG and connected through a cross-attention layer.

Official checkpoint	Role	Inputs and outputs	Legacy context fields	License and status
`Qwen/Qwen-VL`	Pretrained Base model; not instruction-aligned chat	Image, text, and bounding-box input; text and bounding-box-style text output	`seq_length: 2048`; `max_position_embeddings: 8192`	Tongyi Qianwen License; public weights
`Qwen/Qwen-VL-Chat`	Aligned multimodal assistant	Image and text conversation, multiple images, multi-turn chat, and grounding	`seq_length: 2048`; `max_position_embeddings: 8192`	Tongyi Qianwen License; public weights

The checkpoints are often described as “7B” because their language backbone was initialized from Qwen-7B. Use the exact repository ID instead of treating the label as a precise download-size or memory requirement.

Context: 2,048 and 8,192 are different fields

The shipped legacy configuration declares a sequence length of 2,048 and maximum position embeddings of 8,192. The official repository also published profiling scenarios extending through 8,192 generated-token positions with one image. These facts should not be collapsed into an unconditional “8K prompt window” claim.

For compatibility work, treat 2,048 as the conservative configured sequence length unless the exact pinned repository revision, runtime, long-context behavior, image-token cost, and generation budget have been validated together. An image consumed 258 tokens in the Qwen Team’s historical profiling setup, and practical capacity changes with the input and runtime.

What the original Qwen-VL models support

Image captioning and visual question answering.
Reading English and Chinese text in images, including document and chart questions.
Multiple images interleaved with text in Qwen-VL-Chat conversations.
Visual grounding that represents regions as coordinates in generated text markup.
Multi-turn image conversations through the model’s custom chat interface.

The original Qwen-VL checkpoints do not natively support video or audio. Video examples, Qwen3-VL utilities, and current multimodal hosted IDs belong to later products and must not be inserted into this page as if they were original Qwen-VL features.

Official historical Qwen-VL-Chat setup

The following code follows the official Qwen-VL-Chat model card. It uses the checkpoint’s custom repository code, which is why trust_remote_code=True is required.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Qwen/Qwen-VL-Chat"

tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    trust_remote_code=True,
)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    trust_remote_code=True,
).eval()

query = tokenizer.from_list_format(
    [
        {
            "image": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg"
        },
        {"text": "Describe this image."},
    ]
)

response, history = model.chat(
    tokenizer,
    query=query,
    history=None,
)
print(response)

This documentation-derived example was not executed for this update, so no output is claimed. In production, pin and inspect the repository revision before allowing remote custom code to execute. The official historical requirements list Python 3.8+, PyTorch 1.12+ with 2.0+ recommended, and CUDA 11.4+ recommended for GPU users; compatibility with a newer environment must be tested rather than assumed.

License boundaries

The original Qwen-VL repository’s license is the Tongyi Qianwen License Agreement dated August 23, 2023. It is not Apache 2.0. The agreement grants use, modification, and distribution rights subject to conditions; it also contains a commercial-use threshold above 100 million monthly active users and a restriction on using the materials or their output to improve another large language model outside Tongyi Qianwen or its derivatives.

This summary is not legal advice. Read the complete license for the exact repository and obtain legal review for a commercial or redistribution decision.

Limitations of the original 2023 models

No native video or audio input.
Visual input is based on a fixed 448 × 448 encoder design rather than the dynamic-resolution approach introduced later.
Bounding boxes are generated tokens, not guaranteed detections; coordinates must be checked against the source image.
OCR, counting, spatial reasoning, and named-entity recognition can be wrong or incomplete.
The custom-code dependency increases version-pinning and supply-chain review requirements.
Historical benchmark scores do not establish current production quality or safety.

When to use Qwen-VL and when to migrate

Use the original Qwen-VL models for reproducibility, research on the 2023 architecture, or maintenance of a system already tied to Qwen/Qwen-VL or Qwen/Qwen-VL-Chat. For a new project, compare the later Qwen2-VL, Qwen2.5-VL, and Qwen3-VL references, then use the Models hub to identify the current generation.

For official checkpoint files and local options, continue to the Qwen download guide. If you need a hosted service rather than these downloaded weights, choose the exact current provider model in the Qwen API guide and verify its price and region in the Qwen pricing guide. Do not reuse the legacy checkpoint name as a current API ID without provider documentation.

Verification method and test limits

For this update, we checked the official Qwen-VL GitHub repository, the official Qwen-VL and Qwen-VL-Chat Hugging Face repositories, their configuration and license files, and the Qwen-VL paper. We did not download the legacy weights, execute their custom code, or reproduce the historical benchmark, memory, speed, grounding, or OCR results. No site-generated answer or performance figure is claimed.

Frequently asked questions

Is Qwen-VL the same as Qwen2-VL?

No. Qwen-VL and Qwen-VL-Chat are the original 2023 checkpoints. Qwen2-VL is a separate 2024 family with dynamic resolution, video support, M-ROPE, and 2B, 7B, and 72B variants.

Does the original Qwen-VL support video?

No. Its documented native visual input is image-based. Video belongs to later Qwen vision families.

Is Qwen-VL licensed under Apache 2.0?

No. The original checkpoints use the Tongyi Qianwen License Agreement dated August 23, 2023.

Official sources

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

Qwen-VL (2023): Original Models, Setup, and Limits