Qwen3-VL-4B-Instruct

Qwen3-VL-4B-Instruct is a powerful vision-language model with 4 billion parameters, developed by Qwen team at Alibaba Cloud. This model represents the latest generation in the Qwen series, offering superior text understanding, deeper visual perception, extended context length, and enhanced spatial and video comprehension capabilities.

Model Description

Qwen3-VL-4B-Instruct is a multimodal large language model that seamlessly integrates visual and textual understanding. Key capabilities include:

Visual Agent Capabilities: Operates PC/mobile GUIs, recognizes elements, understands functions, invokes tools, and completes tasks
Code Generation from Visuals: Generates Draw.io diagrams, HTML/CSS/JS code from images and videos
Advanced Spatial Reasoning: 2D and 3D grounding for spatial perception and embodied AI applications
OCR Excellence: Supports 32 languages with robust handling of low-light and blurred images
Video Understanding: Full recall and second-level indexing with timestamp-grounded event localization
Extended Context: 256K native context length (expandable to 1M tokens)
STEM/Math Reasoning: Advanced causal analysis and mathematical reasoning capabilities

Repository Contents

This repository contains model files in both SafeTensors and GGUF formats:

File	Size	Format	Description
`qwen3-vl-4b-instruct-abliterated.safetensors`	~8GB	SafeTensors	Main model weights (abliterated version)
`qwen3-vl-4b-instruct-abliterated-f16.gguf`	~8GB	GGUF FP16	Quantized format for efficient inference

Total Repository Size: ~16GB

Note: Model files are currently being downloaded. File sizes are estimates based on the 4B parameter count.

Hardware Requirements

Minimum Requirements

VRAM: 12GB GPU (for FP16 inference)
RAM: 16GB system memory
Disk Space: 20GB available storage

Recommended Requirements

VRAM: 16GB+ GPU (NVIDIA RTX 4070 or better)
RAM: 32GB system memory
Disk Space: 30GB available storage
GPU: CUDA-compatible GPU with compute capability 7.0+

Performance Notes

Enable flash_attention_2 for better acceleration and memory efficiency
Multi-image and video scenarios benefit significantly from flash attention
BF16 precision recommended for optimal quality-performance balance

Usage Examples

Basic Image Understanding

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
from PIL import Image

# Load model and processor
model = Qwen3VLForConditionalGeneration.from_pretrained(
    "E:/huggingface/qwen3-vl-4b-instruct",
    torch_dtype="auto",
    device_map="auto",
    attn_implementation="flash_attention_2"
)
processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-4B-Instruct")

# Prepare image and text input
image = Image.open("path/to/your/image.jpg")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": "Describe this image in detail."}
        ]
    }
]

# Process and generate
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], images=[image], return_tensors="pt").to(model.device)

# Generate response
outputs = model.generate(
    **inputs,
    max_new_tokens=512,
    top_p=0.8,
    temperature=0.7
)
response = processor.batch_decode(outputs, skip_special_tokens=True)[0]
print(response)

Video Understanding

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
import cv2

# Load model
model = Qwen3VLForConditionalGeneration.from_pretrained(
    "E:/huggingface/qwen3-vl-4b-instruct",
    torch_dtype="auto",
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-4B-Instruct")

# Extract video frames
def extract_frames(video_path, num_frames=8):
    cap = cv2.VideoCapture(video_path)
    frames = []
    frame_count = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    indices = [int(i * frame_count / num_frames) for i in range(num_frames)]

    for idx in indices:
        cap.set(cv2.CAP_PROP_POS_FRAMES, idx)
        ret, frame = cap.read()
        if ret:
            frames.append(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
    cap.release()
    return frames

# Process video
frames = extract_frames("path/to/video.mp4")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "video", "video": frames},
            {"type": "text", "text": "What happens in this video? Provide timestamps."}
        ]
    }
]

# Generate analysis
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], videos=[frames], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024)
print(processor.decode(outputs[0], skip_special_tokens=True))

GUI Agent Interaction

from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
from PIL import ImageGrab

# Load model
model = Qwen3VLForConditionalGeneration.from_pretrained(
    "E:/huggingface/qwen3-vl-4b-instruct",
    torch_dtype="auto",
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("Qwen/Qwen3-VL-4B-Instruct")

# Capture screenshot
screenshot = ImageGrab.grab()

# Analyze UI
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": screenshot},
            {"type": "text", "text": "Identify all clickable elements and their functions."}
        ]
    }
]

text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], images=[screenshot], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)
print(processor.decode(outputs[0], skip_special_tokens=True))

Model Specifications

Architecture

Parameters: 4 billion (dense architecture)
Precision: BF16 tensor type
Context Length: 256K tokens (native), expandable to 1M
Vision Encoder: DeepStack multi-level ViT with fine-grained feature fusion
Positional Encoding: Interleaved-MRoPE for temporal, spatial dimensions
Video Processing: Text-Timestamp Alignment for precise event localization

Supported Languages (OCR)

32 languages including English, Chinese, Japanese, Korean, Arabic, French, German, Spanish, and more.

Model Innovations

Interleaved-MRoPE: Full-frequency coverage across temporal, width, and height dimensions
DeepStack: Multi-level ViT feature fusion for enhanced image-text alignment
Text-Timestamp Alignment: Precise timestamp-grounded event localization in videos

Performance Tips

Optimization Strategies

Enable Flash Attention 2: Significantly improves memory efficiency and speed
```
attn_implementation="flash_attention_2"
```

Batch Processing: Process multiple images/videos in batches when possible

inputs = processor(text=texts, images=images, return_tensors="pt", padding=True)

Quantization: Use GGUF format for reduced memory footprint
- FP16: ~8GB (included in repository)
- Consider INT8/INT4 quantization for edge deployment
Context Management: For long contexts, use sliding window or chunking strategies
Generation Parameters:
- Vision-Language Tasks: top_p=0.8, temperature=0.7
- Code Generation: top_p=0.9, temperature=0.3
- Creative Tasks: top_p=0.95, temperature=0.9

Memory Optimization

Use device_map="auto" for automatic multi-GPU distribution
Enable gradient checkpointing for training/fine-tuning
Clear cache between inference runs: torch.cuda.empty_cache()

License

This model is licensed under Apache License 2.0.

Key Points:

✅ Commercial use permitted
✅ Private use permitted
✅ Modification and distribution allowed
⚠️ Must include copyright notice and license
⚠️ Must state significant changes
❌ No trademark use
❌ No warranty provided

Full license text: https://www.apache.org/licenses/LICENSE-2.0

Citation

If you use Qwen3-VL-4B-Instruct in your research or applications, please cite:

@article{qwen3vl2025,
  title={Qwen3-VL: The Next Generation Vision-Language Model},
  author={Qwen Team},
  journal={arXiv preprint},
  year={2025},
  organization={Alibaba Cloud}
}