---
title: Ministral WebGPU
emoji: ⚡️
colorFrom: red
colorTo: yellow
sdk: static
pinned: false
license: apache-2.0
short_description: Frontier multimodal AI, running entirely in your browser.
app_build_command: npm run build
app_file: dist/index.html
models:
- mistralai/Ministral-3-3B-Instruct-2512-ONNX
- mistralai/Ministral-3-3B-Instruct-2512
---
Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

# AI Multimodal WebGPU Assistant

**Developer:** Muhammad Abdullah Rasheed
**Research Assistant @ Cambridge | MSc Data Science & AI '25 | Google WTM Scholar**

## Overview

This project demonstrates cutting-edge browser-based AI by running a complete 3B parameter multimodal language model entirely client-side using WebGPU acceleration. No servers, no API calls, no data sent anywhere - complete privacy and instant inference.

## Key Features

- **Privacy-First Architecture**: The entire Ministral-3B model runs locally in your browser using WebGPU - your video feed never leaves your device
- **Real-Time Multimodal AI**: Live camera feed processing with visual question answering capabilities
- **WebGPU Acceleration**: Leveraging the latest browser GPU APIs for near-native performance
- **Zero Backend Dependencies**: No API keys, no server calls, no external services required
- **Cross-Platform**: Works seamlessly across modern browsers with WebGPU support

## Technical Stack

- **Model**: Ministral-3-3B-Instruct (quantized for browser deployment)
- **Runtime**: Transformers.js for in-browser inference
- **Compute**: WebGPU API for GPU acceleration
- **Frontend**: Modern JavaScript with WebAssembly integration

## Use Cases

- Visual question answering from live camera feed
- Real-time scene understanding and description
- Privacy-sensitive AI applications
- Edge computing demonstrations
- Educational tool for AI and browser technologies

## Why This Matters

This project showcases the future of AI deployment - moving powerful language models from cloud servers to the edge, where they can provide instant, private, and accessible intelligence without compromising user privacy or requiring expensive infrastructure.

## Author

**Muhammad Abdullah Rasheed**  
Research Assistant | AI & Machine Learning Researcher  
- 🎓 MSc Data Science & AI '25, Google WTM Scholar  
- 🔬 Research areas: Computer Vision, NLP, Climate AI  
- 💼 Experience: Gesture Recognition, Backend Development, ML Engineering  
- 🔗 [LinkedIn](https://www.linkedin.com/in/muhammad-abdullahrasheed-/) | [GitHub](https://github.com/Abdullahrasheed45) | [HuggingFace](https://huggingface.co/Abdullahrasheed45)

## License

Apache-2.0