> ## 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.

# Benchmarks

> Performance evaluation of Qwen3-VL models across visual and text-centric tasks

## Overview

Qwen3-VL has been extensively evaluated on a wide range of benchmarks to demonstrate its capabilities in both visual understanding and text processing. Below are the comprehensive benchmark results for all model variants.

## Visual Tasks Performance

### Large Models (235B & 30B MoE)

#### Instruct Models - Visual Tasks

![Qwen3-VL Large Instruct Models - Visual Tasks](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3-VL/table_nothinking_vl.jpg)

**Qwen3-VL-235B-A22B-Instruct** and **Qwen3-VL-30B-A3B-Instruct** demonstrate state-of-the-art performance across visual benchmarks.

#### Thinking Models - Visual Tasks

![Qwen3-VL Large Thinking Models - Visual Tasks](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-VL/table_thinking_vl_.jpg)

**Qwen3-VL-235B-A22B-Thinking** and **Qwen3-VL-30B-A3B-Thinking** show enhanced reasoning capabilities on complex visual tasks.

#### MoE Models Comparison

![Qwen3-VL-30B-A3B Instruct](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3-VL/table_nothinking_vl-30a3.jpg)

![Qwen3-VL-30B-A3B Thinking](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3-VL/table_thinking_vl_30A3.jpg)

### Smaller Models (2B-32B Dense)

#### 2B & 32B Models - Visual Tasks

![Qwen3-VL 2B & 32B Instruct - Visual](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3-VL/qwen3vl_2b_32b_vl_instruct.jpg)

![Qwen3-VL 2B & 32B Thinking - Visual](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3-VL/qwen3vl_2b_32b_vl_thinking.jpg)

Comprehensive comparison of **Qwen3-VL-2B** and **Qwen3-VL-32B** models in both Instruct and Thinking editions across visual understanding benchmarks.

## Text-Centric Tasks Performance

### Large Models (235B & 30B MoE)

#### Instruct Models - Text Tasks

![Qwen3-VL Large Instruct - Text Tasks](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3-VL/table_nothinking_text.jpg)

Qwen3-VL demonstrates **text understanding on par with pure LLMs**, showing seamless text-vision fusion.

#### Thinking Models - Text Tasks

![Qwen3-VL Large Thinking - Text Tasks](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3-VL/table_thinking_text.jpg)

Thinking editions show enhanced performance on reasoning-heavy text tasks.

#### MoE Model - Text Tasks

![Qwen3-VL-30B-A3B Text Performance](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3-VL/table_nothinking_text-30a3.jpg)

### Smaller Models (4B & 8B Dense)

#### 4B & 8B Models - Text Tasks

![Qwen3-VL 4B & 8B Instruct - Text](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3-VL/qwen3vl_4b_8b_text_instruct.jpg)

![Qwen3-VL 4B & 8B Thinking - Text](https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3-VL/qwen3vl_4b_8b_text_thinking.jpg)

Performance comparison of **Qwen3-VL-4B** and **Qwen3-VL-8B** models on text-centric benchmarks.

## Key Capabilities

### Visual Understanding

* **Image Recognition**: State-of-the-art performance on standard vision benchmarks
* **OCR**: Support for 32 languages with robustness to challenging conditions
* **Document Parsing**: Advanced layout understanding and structure extraction
* **Object Grounding**: Precise 2D and 3D object localization
* **Video Understanding**: Long-form video comprehension with temporal reasoning

### Text Processing

* **Pure Text Tasks**: Performance comparable to dedicated LLMs
* **Multimodal Reasoning**: Seamless integration of visual and textual information
* **STEM/Math**: Enhanced reasoning capabilities, especially in Thinking editions
* **Multilingual**: Strong performance across multiple languages

### Specialized Capabilities

* **Spatial Understanding**: Advanced 3D reasoning and spatial relationships
* **Coding**: Visual coding from screenshots to HTML/CSS/JavaScript
* **Agent Tasks**: GUI interaction and tool use
* **Long Context**: Native 256K tokens, expandable to 1M

## Evaluation Settings

To ensure reproducibility, we provide our official evaluation configuration:

### Inference & Evaluation

* **Inference Runtime**: [vLLM](https://github.com/vllm-project/vllm)
* **Evaluation Frameworks**:
  * [VLMEvalKit](https://github.com/open-compass/VLMEvalKit)
  * [lmms-eval](https://github.com/EvolvingLMMs-Lab/lmms-eval)

### Generation Hyperparameters

#### Instruct Models

```bash theme={null}
export greedy='false'
export seed=3407
export top_p=0.8
export top_k=20
export temperature=0.7
export repetition_penalty=1.0
export presence_penalty=1.5
export out_seq_length=32768
```

#### Thinking Models

```bash theme={null}
export greedy='false'
export seed=1234
export top_p=0.95
export top_k=20
export repetition_penalty=1.0
export presence_penalty=0.0
export temperature=0.6
export out_seq_length=40960
```

### Notes on Evaluation

> [!NOTE]
>
> * For certain benchmarks, evaluation prompts were slightly modified for better performance
> * Some benchmarks are internally constructed; reproduction code will be released
> * Detailed methodology will be documented in the technical report

## Benchmark Categories

### Visual Tasks Evaluated

* General vision understanding
* OCR and document analysis
* Object detection and grounding
* Video question answering
* Spatial reasoning
* Visual coding
* Agent tasks (GUI interaction)

### Text-Centric Tasks Evaluated

* Natural language understanding
* Mathematical reasoning
* Code understanding and generation
* Logical reasoning
* Common sense reasoning
* Multilingual comprehension

## Performance Highlights

### Qwen3-VL-235B-A22B

* Best-in-class performance across most benchmarks
* State-of-the-art visual reasoning
* Comparable to pure LLMs on text tasks

### Qwen3-VL-30B-A3B

* Excellent performance-to-cost ratio
* MoE architecture for efficient inference
* Strong across both visual and text tasks

### Qwen3-VL-2B to 32B

* Scalable performance across model sizes
* 2B suitable for edge deployment
* 32B competitive with larger models
* Thinking editions show consistent improvements

## Related Resources

* [Model Cards](https://mintlify.com/QwenLM/Qwen3-VL/resources/model-cards.md) - Download links for all models
* [Quickstart](https://mintlify.com/QwenLM/Qwen3-VL/quickstart.md) - Get started with inference
* [Technical Paper](https://arxiv.org/pdf/2511.21631) - In-depth technical details


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