Qwen3-VL-abliterated-MAX-Fast

Qwen3-VL-abliterated-MAX-Fast is an experimental, high-performance visual reasoning and optical character recognition (OCR) workspace. Powered by the unredacted prithivMLmods/Qwen3-VL-4B-Instruct-Unredacted-MAX architecture, this suite is designed to deliver uncensored, raw image-to-text processing, precise text transcription, and intricate scene understanding. The web application features a standalone, custom-engineered interface built on vanilla web technologies via a headless Gradio framework. It provides a drag-and-drop media drop zone, inline state validation notifications, and a real-time raw output token stream. Optimized to harness active GPU acceleration with Flash Attention 2, it allows developers and security researchers to completely bypass structural system guardrails during complex visual reasoning evaluations.

Key Features

• Abliterated Vision Engine: Utilizes an unredacted version of the Qwen3-VL-4B-Instruct model to secure unrestricted vision-language processing and uncensored textual rendering from image inputs.
• Custom Headless Interface: Houses a sleek, dark terminal-inspired frontend layout designed with embedded JavaScript handling for real-time asset syncing and dynamic UI state response.
• Streaming Token Output: Displays responses step-by-step using text streamer loops that push chunks directly onto the output viewport as they are decoded by the transformer.
• Advanced Pipeline Controls: Offers modular option parameters to manually fine-tune token parameters including Maximum New Tokens, Temperature, Top-p, Top-k, and Repetition Penalty.
• Streamlined Export Utilities: Equipped with built-in instant click tools to quickly copy the entire output text block to the clipboard or download the raw layout response as a clean .txt file.

Repository Structure

├── images/
│   ├── 1.jpg
│   └── 2.jpg
├── app.py
├── LICENSE.txt
├── pre-requirements.txt
├── pyproject.toml
├── README.md
└── requirements.txt

Installation and Requirements

To configure Qwen3-VL-abliterated-MAX-Fast locally, set up an environment with the dependencies listed below. A system containing a modern, CUDA-compatible GPU is required for optimal inference speeds and Flash Attention execution.

Standard PIP Installation

1. Install Pre-requirements Ensure your local package manager is upgraded to align with modern build conditions:

pip install pip>=26.1

2. Install Core Dependencies Install the core deep learning stack, web layers, and document processing utilities:

pip install -r requirements.txt

Running with uv (Recommended)

uv is an ultra-fast Python package and project manager written in Rust, which guarantees rapid virtual environment setup and deterministic dependency syncing.

**Step 1 — Install uv**

• macOS / Linux: curl -LsSf https://astral.sh/uv/install.sh | sh
• Windows: powershell -c "irm https://astral.sh/uv/install.ps1 | iex"

Step 2 — Clone the repository

git clone https://github.com/PRITHIVSAKTHIUR/Qwen3-VL-abliterated-MAX-Fast.git
cd Qwen3-VL-abliterated-MAX-Fast

Step 3 — Initialize the project and install dependencies

uv sync

Step 4 — Run the script

uv run app.py

Core Requirements List

The application builds on the following core dependencies (defined in requirements.txt):

git+https://github.com/huggingface/transformers.git@v4.57.6
git+https://github.com/huggingface/accelerate.git
git+https://github.com/huggingface/peft.git
transformers-stream-generator
huggingface_hub
qwen-vl-utils
sentencepiece
opencv-python
torch==2.11.0
torchvision
matplotlib
pdf2image
requests
pymupdf
kernels
hf_xet
spaces
pillow
gradio==6.15.0
fpdf
timm
av

Usage

Once the FastAPI web layout initiates, load the app locally by pointing your browser to the terminal endpoint (typically http://127.0.0.1:7860/).

1. Upload Asset: Drop an asset sheet, image document, page screenshot, or scene file into the dashed orange uploader area.
2. Write Instruction: Type a descriptive instruction directive inside the Query Input box (e.g., "Read all visible text in the image" or "Describe the image in detail").
3. Advanced Settings (Optional): Tweak advanced sampler configurations like Temperature, Top-p, and Repetition Penalty to adjust output variations.
4. Execute: Click Run Inference to pass parameters to the backend. The results will immediately start streaming into the Raw Output Stream viewport.

License and Source

• License: Apache License 2.0
• GitHub Repository: https://github.com/PRITHIVSAKTHIUR/Qwen3-VL-abliterated-MAX-Fast.git