wangkanai / qwen3-vl-2b-thinking

Qwen3-VL-2B-Thinking (Abliterated)

A 2-billion parameter vision-language model from the Qwen3-VL family, featuring abliterated safety filters for unrestricted generation and enhanced reasoning capabilities. This model combines visual understanding with text generation, enabling multimodal analysis and creative applications.

Model Description

Qwen3-VL-2B-Thinking-Abliterated is a modified version of Qwen3-VL-2B optimized for:

• Vision-Language Understanding : Process images and generate contextual text responses
• Multimodal Reasoning : Analyze visual content with detailed explanations
• Unrestricted Generation : Abliterated safety layers for creative freedom
• Thinking Mode : Enhanced reasoning and step-by-step analysis capabilities
• Efficient Inference : 2B parameters for consumer hardware deployment

Key Features :

• Visual question answering (VQA)
• Image captioning and description
• Visual reasoning and analysis
• Multimodal conversation
• Creative image interpretation

Repository Contents

qwen3-vl-2b-thinking/
├── qwen3-vl-2b-thinking-abliterated.safetensors  # PyTorch model (4.0GB)
└── qwen3-vl-2b-thinking-abliterated.gguf         # GGUF quantized (3.3GB)

Total Repository Size : ~7.3GB

Model Formats

Hardware Requirements

Minimum Requirements

• VRAM : 6GB (GGUF quantized inference)
• RAM : 8GB system memory
• Disk Space : 8GB available
• GPU : CUDA-compatible (RTX 2060+) or Apple Silicon

Recommended Requirements

• VRAM : 8-12GB (SafeTensors full precision)
• RAM : 16GB system memory
• Disk Space : 10GB available
• GPU : RTX 3060 Ti+ or Apple M1 Pro+

Performance Estimates

• GGUF on 8GB VRAM : ~15-25 tokens/sec
• SafeTensors on 12GB VRAM : ~20-35 tokens/sec
• CPU inference : ~2-5 tokens/sec (not recommended)

Usage Examples

Using Transformers (SafeTensors)

from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
from PIL import Image
import torch

# Load model and processor
model_path = "E:/huggingface/qwen3-vl-2b-thinking"
model = Qwen2VLForConditionalGeneration.from_pretrained(
    model_path,
    torch_dtype=torch.float16,
    device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_path)

# Load image
image = Image.open("image.jpg")

# Create conversation
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("cuda")

# Generate response
with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=256)
response = processor.batch_decode(outputs, skip_special_tokens=True)[0]

print(response)

Using llama.cpp (GGUF)

# Download llama.cpp with vision support
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
make

# Run inference with image
./llama-cli \
  --model "E:/huggingface/qwen3-vl-2b-thinking/qwen3-vl-2b-thinking-abliterated.gguf" \
  --image "image.jpg" \
  --prompt "Describe this image:" \
  --n-gpu-layers 32 \
  --ctx-size 4096

Using Ollama (GGUF)

# Create Modelfile
cat > Modelfile <<EOF
FROM E:/huggingface/qwen3-vl-2b-thinking/qwen3-vl-2b-thinking-abliterated.gguf
PARAMETER temperature 0.7
PARAMETER top_p 0.9
EOF

# Create Ollama model
ollama create qwen3-vl-thinking -f Modelfile

# Run interactive session
ollama run qwen3-vl-thinking "Analyze this image: image.jpg"

Visual Question Answering

# Detailed analysis with thinking mode
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "image": image},
            {"type": "text", "text": "Think step-by-step and explain what's happening in this image."}
        ]
    }
]

# Model will provide detailed reasoning in its response

Model Specifications

Architecture

• Model Type : Vision-Language Transformer
• Base Architecture : Qwen3-VL
• Parameters : 2 billion
• Modifications : Abliterated safety layers, enhanced reasoning
• Vision Encoder : ViT-based image encoder
• Text Decoder : Qwen3 transformer decoder

Technical Details

• Precision : FP16 (SafeTensors), Quantized (GGUF)
• Context Length : 4096 tokens
• Image Resolution : 448x448 (default), up to 1024x1024
• Vocabulary Size : ~151,000 tokens
• Training : Multimodal pretraining + instruction tuning

Supported Tasks

• Image captioning
• Visual question answering
• Scene understanding
• Object detection (descriptive)
• Visual reasoning
• Image-to-text generation
• Multimodal conversation

Performance Tips

Optimization Strategies

1. VRAM Optimization :

   # Use 8-bit quantization
   model = Qwen2VLForConditionalGeneration.from_pretrained(
       model_path,
       load_in_8bit=True,
       device_map="auto"
   )
   

2. Image Preprocessing :

   # Resize large images
   from PIL import Image
   image = Image.open("large_image.jpg")
   image = image.resize((448, 448))
   

3. Batch Processing :

   # Process multiple images efficiently
   images = [Image.open(f"image{i}.jpg") for i in range(4)]
   inputs = processor(images=images, text=prompts, return_tensors="pt")
   

4. GGUF Performance :

   • Use --n-gpu-layers 32 for GPU acceleration
   • Adjust --ctx-size based on available VRAM
   • Use --threads for CPU optimization

Generation Parameters

generation_config = {
    "max_new_tokens": 256,
    "temperature": 0.7,
    "top_p": 0.9,
    "do_sample": True,
    "repetition_penalty": 1.1
}

outputs = model.generate(**inputs, **generation_config)

Abliteration Notice

This model has been abliterated (safety filters removed) for:

• Unrestricted creative content generation
• Research and experimentation
• Artistic applications without content restrictions

Important : Users are responsible for ethical use and compliance with local laws. This model may generate unrestricted content.

License

Licensed under Apache 2.0 . Free for commercial and research use with attribution.

Key provisions:

• ✅ Commercial use permitted
• ✅ Modification and distribution allowed
• ✅ Private use permitted
• ⚠️ Provide attribution and license notice
• ⚠️ State changes if modified

Full license: Apache License 2.0

Citation

@misc{qwen3vl2b-thinking-abliterated,
  title={Qwen3-VL-2B-Thinking-Abliterated},
  author={Qwen Team and Community Contributors},
  year={2025},
  howpublished={\url{https://huggingface.co/Qwen}},
  note={Abliterated vision-language model with enhanced reasoning}
}

Official Resources

• Qwen Official : https://github.com/QwenLM/Qwen
• Transformers Docs : https://huggingface.co/docs/transformers
• llama.cpp : https://github.com/ggerganov/llama.cpp
• Model Family : https://huggingface.co/Qwen

Acknowledgments

• Qwen Team : Original Qwen3-VL architecture and pretraining
• Community : Abliteration techniques and reasoning enhancements
• Hugging Face : Model hosting and transformers library