Citation and evidence

QvQ

Report a problem with this article

Edit article
More

Use this article

Raw MarkdownExplore connections

Improve this page

Revision historyDiscuss this article

Topics

Chinese AIMultimodal AIReasoning Models

Cite this article

QvQ (styled QVQ) is a family of experimental visual reasoning models from the Qwen team at Alibaba. The models extend the large language model reasoning approach to images: instead of answering a question about a picture in a single step, a QVQ model produces a long internal chain of thought and reasons about the visual content before giving a final answer. The first release, QVQ-72B-Preview, appeared on 25 December 2024 as the visual counterpart to the text-only reasoning model QwQ (QwQ-32B-Preview), which the team had published the previous month. [1][2] A second, redesigned model, QVQ-Max, was announced on 28 March 2025. [3]

Qwen logo
Qwen logo. w:Alibaba Cloud, via Wikimedia Commons, Apache License 2.0.

The name follows the team's habit of pairing a "Q" for Qwen with a playful repetition: where QwQ is the text reasoner, QVQ adds a "V" for vision. [1]

Relationship to QwQ and Qwen2-VL

QVQ-72B-Preview is built on Qwen2-VL-72B, Alibaba's open-weight multimodal model, rather than being trained from scratch. The Hugging Face model card lists the base model as Qwen/Qwen2-VL-72B and tags the model for image-text-to-text use. [4] In effect QVQ takes the perception stack of Qwen2-VL and layers a long reasoning behaviour on top, the same idea QwQ-32B-Preview applied to text-only problems. [1][2] The team described both QwQ and QVQ as exploratory research releases rather than finished products. [1]

QVQ is positioned as a reasoning_model for the multimodal setting. The Qwen blog framed the goal in terms of combining language and vision so a model can "see the world with wisdom," and likened the intended behaviour to a person who works through a hard physics diagram step by step instead of guessing. [1]

QVQ-72B-Preview

QVQ-72B-Preview is the 72-billion-parameter preview model released on 25 December 2024 under the title "QVQ: To See the World with Wisdom." [1] It was published with open weights on Hugging Face alongside an online demo space, and shares the repository lineage of Qwen2-VL. [1][4]

The model has several practical constraints that the developers stated plainly. It supports only single-round dialogue and image inputs: it does not accept video, and it is not designed for multi-turn conversation. [4] In day-to-day recognition tasks such as identifying people, animals, or plants, the team reported that QVQ does not improve meaningfully over Qwen2-VL-72B-Instruct, so it is not a drop-in replacement for the base model. [4]

Benchmarks

The Qwen team evaluated QVQ-72B-Preview on four multimodal reasoning benchmarks: MMMU (a university-level, multidisciplinary multimodal exam), MathVista, MathVision, and OlympiadBench. The headline result is a score of 70.3 on MMMU, with larger gains on the math-heavy sets relative to the Qwen2-VL-72B-Instruct baseline. The team compared the preview against OpenAI's o1 (the 17 December 2024 build), GPT-4o (13 May 2024), and Claude 3.5 Sonnet (22 October 2024). [1][4]

BenchmarkQVQ-72B-Previewo1-2024-12-17GPT-4o-2024-05-13Claude 3.5 Sonnet-20241022Qwen2-VL-72B-Instruct
MMMU (val)70.377.369.170.464.5
MathVista (mini)71.471.063.865.370.5
MathVision (full)35.9N/A30.435.625.9
OlympiadBench20.4N/A25.9N/A11.2

Expanded article table

Source: Qwen blog and the QVQ-72B-Preview Hugging Face card. [1][4] On MathVista the preview slightly edged o1 in the reported figures, while on MMMU it trailed o1 but landed close to Claude 3.5 Sonnet and ahead of GPT-4o. The team characterised the overall result as closing much of the gap with leading reasoning systems on these visual tasks. [1]

Limitations noted by the developers

The Qwen team published an unusually candid list of weaknesses for the preview, grouped under four headings. [1][4]

Language mixing and code-switching: the model can occasionally mix languages or switch between them, which can hurt the clarity of an answer. [4]

Recursive reasoning loops: the long chain of thought can fall into circular patterns, producing very long responses that sometimes never reach a final answer. [4]

Safety and ethical considerations: the team stated that robust safety measures are still needed and advised caution when deploying the model. [4]

Performance and benchmark limits: QVQ does not fully replace Qwen2-VL-72B-Instruct, and during multi-step visual reasoning it may gradually lose track of the image content and hallucinate. It also shows no significant gain over the base model on basic recognition. [4]

These caveats, together with the "Preview" label, frame QVQ-72B as a research artifact meant to demonstrate visual chain-of-thought rather than a production system. [1]

Licensing

QVQ-72B-Preview is released under the Qwen license. The Hugging Face metadata records the license as "other" with the license name "qwen" and a link to the model's LICENSE file, the same custom community license Alibaba uses for its larger Qwen weights. [4] The weights, an online demo, and the model card are hosted on Hugging Face. [1][4]

QVQ-Max

On 28 March 2025 the Qwen team announced QVQ-Max under the title "QVQ-Max: Think with Evidence," describing it as the first version of a redesigned visual reasoning model that supersedes the December preview. [3] Unlike the preview, QVQ-Max accepts both images and video, and it is offered through the Qwen Chat interface, where users upload an image or video, ask a question, and press a "Thinking" button to watch the model reason before it answers. [3]

The blog presented QVQ-Max less as a benchmark leader and more as a practical tool, with worked examples spanning visual math problems, charts and diagrams, everyday questions, code, and creative tasks such as helping with illustration or design. [3] One quantitative point the team highlighted is that accuracy on MathVision rises as the maximum length of the model's thinking process is increased, an effect analogous to test-time scaling in text reasoners. [3] The post did not publish a full comparison table of the kind released for the 72B preview. [3]

For the future of the line, the team listed three directions: improving observational accuracy through grounding (so the model can verify the visual evidence behind a claim), building visual agents that carry out multi-step tasks such as operating a phone or computer or playing games, and broadening interaction beyond text. [3] The earlier QVQ-72B-Preview weights remain available on Hugging Face, while QVQ-Max has been distributed through Qwen Chat and Alibaba Cloud's interface rather than as a clearly documented open-weight checkpoint. [3][4]

QVQ sits alongside Alibaba's broader multimodal line, including Qwen2.5-VL, which the team continued to develop in parallel as the general-purpose vision-language family while QVQ pursued the narrower goal of long-form visual reasoning. [3]

Continue reading

More to read

Topics:Chinese AIMultimodal AIReasoning Models

References

  1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13"QVQ: To See the World with Wisdom," Qwen (Alibaba Cloud) blog, 25 December 2024. qwenlm.github.io/...qvq-72b-preview
  2. ^1 ^2"QwQ: Reflect Deeply on the Boundaries of the Unknown," Qwen (Alibaba Cloud) blog. qwenlm.github.io/...qwq-32b-preview
  3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9"QVQ-Max: Think with Evidence," Qwen (Alibaba Cloud) blog, 28 March 2025. qwenlm.github.io/...qvq-max-preview
  4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14"Qwen/QVQ-72B-Preview," Hugging Face model card. huggingface.co/...QVQ-72B-Preview

Improve this article

Add missing citations, update stale details, or write a clearer explanation. Anyone can edit: changes go live when you publish, and every edit is checked by an editor.

6 revisions · v6 · 1,158 words · full history

Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify

Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here

Cite this page: AI Wiki. "QvQ." aiwiki.ai, updated 1 Nov 2026. CC BY 4.0. https://aiwiki.ai/wiki/qvq

Edit article