🧠 BrowserGemma

Gemma 4 multimodal AI running 100% in your browser. Private, free, no server, no API.

✨ What is this?

BrowserGemma is a demo of Google's Gemma 4 multimodal AI (text + vision) running entirely client-side in your browser via WebGPU (with WASM fallback). No data ever leaves your device. No API keys. No server.

The entire frontend is hosted on the Akash Network — a decentralized cloud. This means:

• 🔒 Privacy-first: All inference happens on YOUR device
• 🌐 Decentralized: Hosted on the Akash supercloud, not Big Tech
• 💰 Free to use: No API costs, no accounts
• ⚡ Fast: WebGPU gives 10-100x speedup over WASM on supported browsers
• 📱 Installable: Add to home screen as a PWA

🚀 Try It

Note: First load downloads the model (~hundreds of MB), then it's cached by your browser. Requires Chrome/Edge (latest) for WebGPU. Falls back to WASM on other browsers.

📸 What can it do?

• 🖼️ Image description — Upload any image and ask about it
• 💬 Multimodal chat — Multi-turn conversation with images
• 🎨 Art analysis — Identify styles, emotions, techniques
• 📊 Chart reading — Extract data from visualizations
• 😂 Meme explanation — Upload memes and get AI humor analysis

🛠️ Tech Stack

🏗️ Architecture

Your Browser                    Akash Network
┌──────────────────────┐       ┌──────────────┐
│  React UI            │       │  Nginx       │
│  ┌────────────────┐  │       │  (static     │
│  │ transformers.js│  │◄──────│   HTML/JS/   │
│  │ ┌────────────┐ │  │ HTTP  │   WASM)      │
│  │ │ Gemma 4    │ │  │       └──────────────┘
│  │ │ (WebGPU/   │ │  │
│  │ │  WASM)     │ │  │       HuggingFace CDN
│  │ └────────────┘ │  │       ┌──────────────┐
│  └────────────────┘  │◄──────│  ONNX Model  │
│                      │ HTTPS │  Weights      │
│  🔒 Zero data leaves │       └──────────────┘
│     your device      │
└──────────────────────┘

Key insight: The Akash server only serves static files. ALL AI inference runs in your browser. Model weights are downloaded directly from HuggingFace's CDN and cached locally.

📦 Local Development

# Clone
git clone https://github.com/toxmon/browsrgemma.git
cd browsrgemma

# Install
npm install

# Dev server
npm run dev

# Production build
npm run build

🐳 Docker Build

docker build -t toxmon/browsrgemma:0.1.0 .

☁️ Deploy on Akash

Option 1: One-Click (via Cloudmos Console)

1. Go to console.akash.network
2. Click "Deploy"
3. Paste the deploy.yml contents
4. Fund with AKT and deploy

Option 2: CLI

# Install Akash CLI (if needed)
# See: https://akash.network/docs/getting-started/cli/

# Create deployment
akash tx deployment create deploy.yml --from $KEY --chain-id akashnet-2 --fees 5000uakt

# View bids
akash lease list

# Send manifest to provider
akash provider send-manifest deploy.yml <LEASE-ID>

Option 3: Build & Push Your Own Image

# Build
docker build -t your-dockerhub-user/browsrgemma:0.1.0 .

# Push
docker push your-dockerhub-user/browsrgemma:0.1.0

# Update deploy.yml image field, then deploy

💰 Cost

Akash hosting for a static site: ~$0.01-0.05/month in AKT. Essentially free.

📱 PWA

BrowserGemma is a Progressive Web App. On mobile:

1. Open the site in Chrome/Safari
2. Tap "Add to Home Screen"
3. It launches like a native app

🎯 Why This Matters

• Privacy: Your images and text never leave your device. No server logs, no data collection.
• Accessibility: AI that works offline after first load, no API keys needed.
• Decentralization: Proves that AI demos don't need centralized hosting.
• Performance: WebGPU enables real-time inference that was impossible in browsers a year ago.

🙏 Credits

• Google DeepMind — Gemma 4 model
• Hugging Face — transformers.js, ONNX conversion
• Akash Network — Decentralized cloud hosting

📄 License

MIT

---