QVQ Max

Visual reasoning model with always-on chain-of-thought — solves math diagrams, charts, and complex visual problems step-by-step.

~27.59s
~$0.009
Playground
APIPricing

Chat

0 messages

Press Enter to send, Shift + Enter for new line • Max 5 files (10MB each)

QVQ Max — Visual Reasoning Model API

What is QVQ Max?

QVQ Max is Alibaba Cloud's flagship visual reasoning model, developed by the Qwen team as the production successor to QVQ-72B-Preview. It is purpose-built for tasks that require extended logical reasoning over visual information — not just recognizing what is in an image, but thinking through it step-by-step to arrive at a well-reasoned answer.

Unlike general vision-language models that return instant direct answers, QVQ Max is a thinking-only model: every response begins with a transparent chain-of-thought reasoning process, visible in real time via streaming output. This makes it uniquely suited for complex visual problem-solving in education, research, analytics, and enterprise AI applications where accuracy and explainability are critical.

The model accepts text, images, and video frames within a 131,072-token context window, and always outputs its reasoning chain alongside the final answer.

Key Features

  • •Always-on chain-of-thought reasoning — the model thinks through every problem before answering; reasoning is streamed in real time
  • •131K token context window — supports long multi-image prompts, video frames, and complex visual documents
  • •Multi-image support — compare, contrast, and reason across multiple images in a single call
  • •Video understanding — extract sequential reasoning from video frames with temporal grounding
  • •Streaming-only output — responses are delivered progressively, ideal for interactive and real-time applications
  • •High accuracy on STEM problems — excels at mathematics, physics, and chart interpretation tasks requiring inference, not just recognition

Best Use Cases

STEM Education and Tutoring: Solve math, physics, and chemistry problems from textbook diagrams, handwritten notes, or whiteboard photos — with every reasoning step shown.

Data Analysis and Business Intelligence: Interpret complex charts, financial dashboards, and multi-variable graphs with detailed written explanations of trends and anomalies.

Scientific and Medical Research: Analyze annotated figures, experimental data plots, pathology slides, or research paper diagrams requiring domain-aware reasoning.

Engineering and Architecture Review: Read circuit schematics, system architecture diagrams, or CAD-style drawings and reason about design decisions or implementation steps.

Visual QA for Enterprise: Audit technical documents, annotated maps, compliance screenshots, or instructional manuals where inference over visual content is required.

Code and UI Analysis: Analyze screenshots of interfaces, error outputs, or code diagrams and reason about bugs, design improvements, or implementation steps.

Prompt Tips and Output Quality

For best results with QVQ Max, frame your prompts as explicit questions that require reasoning, not just description. Instead of asking what is in this image, ask what mathematical principle the diagram demonstrates and to solve the problem step by step. The model performs best when it knows what conclusion to reason toward.

Since QVQ Max always shows its reasoning chain, you can reference intermediate steps in follow-up prompts. This makes it ideal for interactive, multi-turn analytical sessions.

For video inputs, specify the time range or frame numbers you want analyzed. For multi-image comparisons, number your images in the prompt and ask for explicit comparisons.

Be aware that thinking tokens are billed as output — for simple tasks, consider Qwen3 VL Flash as a more cost-effective alternative.

FAQs

Is the chain-of-thought reasoning optional? No — QVQ Max is a thinking-only model. Reasoning is always generated and cannot be disabled. This is by design, ensuring high accuracy on complex tasks.

Does QVQ Max support video inputs? Yes — it can process video frames within its 131K context window, with temporal reasoning to understand sequences and events over time.

How does QVQ Max compare to Qwen3 VL Flash? QVQ Max is optimized for deep, step-by-step reasoning on complex tasks and is always in thinking mode. Qwen3 VL Flash offers optional thinking mode at significantly lower cost, making it better for high-volume or simpler visual tasks.

Why is QVQ Max streaming-only? Extended reasoning chains make streaming the natural output method — you see the model's thinking as it happens, which is valuable for interactive applications and educational use cases.

What image formats are supported? QVQ Max accepts URLs and base64-encoded images in PNG, JPEG, and WebP formats. Multiple images can be included in a single prompt.

Is QVQ Max suitable for real-time production workloads? It depends on the use case. For tasks requiring deep reasoning accuracy, QVQ Max is excellent. For high-throughput batch processing, Qwen3 VL Flash is more appropriate.