QwenLM drops QVQ-Max for visual reasoning

Most multimodal models are like careless students scanning a test: they can read the labels on a diagram just fine, but the second you ask them to solve the geometry problem underneath, they guess wildly.

Following its experimental QVQ-72B-Preview last December—which the team admitted had plenty of flaws—the Qwen team officially released the first version of QVQ-Max, a dedicated visual reasoning model designed to think through visual evidence step by step.

Thinking longer boosts visual accuracy

The headline takeaway from QVQ-Max is that "test-time compute" works on visual inputs.

On MathVision, a benchmark packed with challenging multimodal math problems, the team showed that simply extending the model’s maximum thinking length leads to continuous improvements in accuracy. Instead of rushing to spit out an answer, the model spends more tokens reasoning through shapes, spatial relationships, and formulas embedded in the images.

Beyond still pictures and complex charts, QVQ-Max can also parse video clips to predict what might happen next based on a visual scene. The team showcased practical applications ranging from students solving diagram-heavy physics equations to office workers breaking down data charts, writing code, or turning rough visual sketches into finished creative scripts.

The roadmap to screen-controlling agents

The Qwen team was candid that this release is strictly a v1, with several critical upgrades already mapped out.

Immediate priorities include grounding techniques to verify that the model's visual observations actually match reality, cutting down on hallucinated details.

More ambitiously, the team plans to evolve QVQ-Max into a full visual agent capable of multi-step computer use—navigating smartphone interfaces, using desktop software, and even playing video games, backed by external tool verification and visual generation.

Why it matters

We think QVQ-Max represents a crucial shift in multimodal AI: moving from passive recognition to active, deliberate visual reasoning.

Reasoning models like OpenAI's o1 proved that giving text models time to ponder dramatically improves logic. QVQ-Max brings that same playbook to pixels. For developers and knowledge workers dealing with schematics, floor plans, and complex dashboards, this is fundamentally more useful than a chatbot that merely describes what it sees.

That said, the approach comes with clear trade-offs. Visual reasoning with extended thinking paths is computationally expensive and slow, making real-time interactions tough. Furthermore, without the promised grounding fixes, any initial misinterpretation of a visual detail will corrupt the entire chain of thought. It is well worth testing today for complex homework and diagram parsing, but relying on it for autonomous screen tasks will have to wait for future iterations.