> ## Documentation Index
> Fetch the complete documentation index at: https://mintlify.com/QwenLM/Qwen3-VL/llms.txt
> Use this file to discover all available pages before exploring further.

# Key Features

> Comprehensive overview of Qwen3-VL's capabilities including visual agents, spatial perception, video understanding, and more

Qwen3-VL represents a comprehensive upgrade in vision-language capabilities, delivering state-of-the-art performance across multiple domains.

## Visual Agent Capabilities

### GUI Interaction

Qwen3-VL can operate both PC and mobile graphical user interfaces:

* **Element Recognition**: Identifies UI components (buttons, forms, menus)
* **Function Understanding**: Comprehends the purpose of interface elements
* **Tool Invocation**: Executes commands and interactions
* **Task Completion**: Performs multi-step workflows autonomously

```python Mobile Agent theme={null}
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "screenshot.png"},
            {"type": "text", "text": "Click the settings button"}
        ]
    }
]
```

```python Computer Use theme={null}
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "desktop.png"},
            {"type": "text", "text": "Open the browser and search for AI news"}
        ]
    }
]
```

### Visual Coding

Generate code from visual inputs:

* **Draw\.io diagrams** → Structured diagram definitions
* **Screenshots** → HTML/CSS/JavaScript code
* **Design mockups** → Frontend implementations
* **Video tutorials** → Code reproduction

## Spatial Understanding

### 2D Grounding

**Precise object localization with multiple output formats:**

* Bounding boxes with relative coordinates
* Point annotations
* Multiple objects simultaneously
* Diverse positioning and labeling tasks

```python theme={null}
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "scene.jpg"},
            {"type": "text", "text": "Locate all people in this image"}
        ]
    }
]
# Returns: [<box>(x1,y1,x2,y2)</box>, ...]
```

### 3D Grounding

**Spatial reasoning in three dimensions:**

* Indoor and outdoor object detection
* 3D bounding box generation
* Viewpoint understanding
* Occlusion reasoning
* Depth perception

> [!TIP]
>
> 3D grounding enables applications in robotics, embodied AI, and augmented reality scenarios.

### Advanced Spatial Perception

* **Position Judgment**: Determines object locations relative to each other
* **Viewpoint Analysis**: Understands camera angles and perspectives
* **Occlusion Detection**: Identifies when objects block others
* **Spatial Reasoning**: Infers relationships between objects in 3D space

## Long Context & Video Understanding

### Context Window

* **Native**: 256K tokens
* **Extended** (with YaRN): Up to 1M tokens
* **Applications**:
  * Full book analysis
  * Hours-long video processing
  * Complete codebase comprehension

### Video Processing

**Comprehensive video understanding capabilities:**

* **Temporal Modeling**: Precise event timing and sequencing
* **Second-level Indexing**: Frame-accurate content retrieval
* **Long Video Support**: Process videos of any length
* **Full Recall**: Maintain context across entire video duration

```python theme={null}
messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "video",
                "video": "lecture.mp4",
                "max_pixels": 256 * 32 * 32,
                "fps": 2.0
            },
            {"type": "text", "text": "Summarize this hour-long lecture"}
        ]
    }
]
```

### Video Features

* **Video OCR**: Extract text from video frames
* **Video Grounding**: Locate specific moments
* **Action Recognition**: Identify activities and events
* **Scene Understanding**: Comprehend context changes

## Enhanced OCR Capabilities

### Language Support

**32 languages** (expanded from 10 in previous versions)

* Latin scripts
* Chinese (Simplified & Traditional)
* Japanese, Korean
* Arabic, Hebrew
* Cyrillic scripts
* And 25+ more

### Robust Text Recognition

**Handles challenging conditions:**

* **Low light**: Dimly lit scenes
* **Blur**: Motion or focus blur
* **Tilt**: Rotated or skewed text
* **Rare characters**: Ancient scripts, specialized glyphs
* **Technical jargon**: Domain-specific terminology

### Document Understanding

* Layout position information
* Structure parsing for long documents
* Table extraction
* Form understanding
* Qwen HTML format output

```python General OCR theme={null}
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "document.pdf"},
            {"type": "text", "text": "Extract all text preserving layout"}
        ]
    }
]
```

```python Key Information Extraction theme={null}
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "invoice.jpg"},
            {"type": "text", "text": "Extract total amount, date, and vendor name"}
        ]
    }
]
```

## Multimodal Reasoning

### Enhanced STEM Performance

**Excels in scientific and mathematical tasks:**

* **Visual Math**: Solve problems from diagrams and equations
* **Causal Analysis**: Identify cause-effect relationships
* **Logical Reasoning**: Evidence-based conclusions
* **Multi-step Problems**: Complex problem decomposition

### Reasoning Capabilities

* Mathematical equation solving
* Scientific diagram interpretation
* Chart and graph analysis
* Statistical reasoning
* Geometry and spatial problems

## Omni Recognition

### Visual Recognition Scope

**Broader, higher-quality pretraining enables "recognize everything":**

> **People & Characters**
>
> * Celebrities
> * Anime characters
> * Historical figures

> **Objects & Products**
>
> * Vehicles (cars, planes)
> * Consumer products
> * Brand recognition

> **Nature & Places**
>
> * Landmarks
> * Flora and fauna
> * Geographical features

### Recognition Features

* **Fine-grained Classification**: Distinguish between similar objects
* **Attribute Recognition**: Identify colors, materials, styles
* **Scene Understanding**: Comprehend overall context
* **Multi-object Recognition**: Handle complex scenes

## Text Understanding

### On Par with Pure LLMs

**Seamless text-vision fusion:**

* No degradation in pure text tasks
* Unified comprehension across modalities
* Lossless information processing
* Consistent performance whether input is text, image, or both

### Applications

* Mixed text and image documents
* Code with visual diagrams
* Scientific papers with figures
* Interleaved content processing

## Performance Highlights

### Visual Tasks

State-of-the-art performance on:

* General visual question answering
* Document VQA
* Chart/diagram understanding
* Scene text recognition
* Multi-image reasoning

### Text-Centric Tasks

Competitive with specialized text models on:

* Language understanding
* Code generation
* Mathematical reasoning
* Knowledge question answering

> [!NOTE]
>
> For detailed benchmark results, see the [Performance](https://mintlify.com/QwenLM/Qwen3-VL/performance) section.

## Accessibility Features

### Cookbook Examples

Explore hands-on examples for each capability:

* [Omni Recognition](https://github.com/QwenLM/Qwen3-VL/blob/main/cookbooks/omni_recognition.ipynb)
* [Document Parsing](https://github.com/QwenLM/Qwen3-VL/blob/main/cookbooks/document_parsing.ipynb)
* [2D Grounding](https://github.com/QwenLM/Qwen3-VL/blob/main/cookbooks/2d_grounding.ipynb)
* [3D Grounding](https://github.com/QwenLM/Qwen3-VL/blob/main/cookbooks/3d_grounding.ipynb)
* [Video Understanding](https://github.com/QwenLM/Qwen3-VL/blob/main/cookbooks/video_understanding.ipynb)
* [Mobile Agent](https://github.com/QwenLM/Qwen3-VL/blob/main/cookbooks/mobile_agent.ipynb)
* [Computer Use](https://github.com/QwenLM/Qwen3-VL/blob/main/cookbooks/computer_use.ipynb)
* [Spatial Understanding](https://github.com/QwenLM/Qwen3-VL/blob/main/cookbooks/spatial_understanding.ipynb)

> [!TIP]
>
> All cookbooks are available as Jupyter notebooks with Google Colab badges for easy experimentation.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.