Chrome's Built-in AI is an experimental feature that integrates Gemini Nano — Google's lightweight large language model — directly into the Chrome browser. Unlike cloud-based AI services, Gemini Nano runs entirely on your device, meaning no data is sent to external servers. This enables offline AI capabilities, enhanced privacy, and lower latency. Gemini Nano is optimized for consumer hardware and is part of Google's broader initiative to bring on-device AI to the web platform via standardized APIs like window.ai.

Step-by-step setup:
1. Download Chrome Canary (version 127+) from the official Chrome channels.
2. Navigate to chrome://flags/#prompt-api-for-gemini-nano and set it to Enabled.
3. Go to chrome://flags/#optimization-guide-on-device-model and set it to Enabled.
4. Restart Chrome. The browser will automatically download the Gemini Nano model (~1.5-2 GB) in the background.
5. Check chrome://on-device-internals/ to monitor the model download status.
6. Once downloaded, visit any page using the window.ai API — the model will be ready for use.
Note: This is an experimental feature and may change. Hardware requirements include at least 8GB RAM for optimal performance.

Gemini Nano supports a wide range of natural language processing tasks through the window.ai API, including:
• Text Generation — Create articles, stories, emails, and creative content.
• Summarization — Condense long articles, documents, or conversations.
• Translation — Translate between dozens of languages with high accuracy.
• Code Generation — Generate and explain code snippets across programming languages.
• Question Answering — Answer factual questions based on provided context.
• Rewriting & Paraphrasing — Improve clarity, tone, or style of existing text.
• Classification & Sentiment Analysis — Analyze text for categories or emotional tone.
All tasks run locally, making it ideal for privacy-sensitive applications.

Yes, Gemini Nano runs completely offline after the initial model download. All inference happens on your device's CPU or GPU — no data is sent to Google servers. This makes it an excellent choice for:
• Handling sensitive documents (legal, medical, financial)
• Working in environments with limited or no internet connectivity
• Applications requiring GDPR/HIPAA compliance
• Enterprise use cases where data must never leave the device
The model weights are stored locally in Chrome's cache. You can verify no network activity during inference using Chrome DevTools. This is a fundamental shift from cloud-based AI services like ChatGPT or Claude.

While Gemini Nano is impressive for an on-device model, it has some limitations:
• Model Size — Smaller than cloud models (GPT-4, Claude 3), so may struggle with highly complex reasoning or very long contexts.
• Context Window — Limited context length (typically 4K-8K tokens) compared to cloud models with 128K+ tokens.
• Hardware Dependent — Performance varies based on your device's RAM and processor. Older devices may experience slower inference.
• No Internet Access — Cannot browse the web or access real-time information.
• Experimental API — The window.ai API is still evolving and subject to change.
• Model Freshness — The model is tied to Chrome versions; updates require browser updates.
For most everyday tasks — summarization, translation, basic coding, content generation — Gemini Nano performs excellently while keeping your data private.

Using window.ai in your web app is straightforward:

1. Check API availability:
if ('ai' in window && window.ai.languageModel) { /* API available */ }

2. Check capabilities:
const caps = await window.ai.languageModel.capabilities();
// caps.available: 'readily' | 'after-download' | 'no'

3. Create a session:
const session = await window.ai.languageModel.create({
  systemPrompt: 'You are helpful.',
  temperature: 0.7,
  topK: 40,
  maxTokens: 512
});

4. Prompt the model:
const result = await session.prompt('Hello!');
Or use streaming: session.promptStreaming('Hello!')

5. Clean up:
session.destroy();

Always wrap calls in try-catch blocks and provide fallback behavior for browsers without API support.

Currently, the Chrome Built-in AI API (Gemini Nano) is only available in:
• Chrome Canary 127+ — with experimental flags enabled
• Chrome Dev — may work but Canary is recommended
• Chromium-based browsers — theoretically possible if they expose the API

Not supported in: Firefox, Safari, Edge (standard), or stable Chrome releases. The API is expected to roll out to stable Chrome in 2025 after sufficient testing. Other browser vendors have shown interest in on-device AI but have not announced concrete plans. This playground includes a simulation mode for unsupported browsers, allowing you to explore the API's capabilities and generate code snippets regardless.

Cost comparison:
• Gemini Nano (on-device): $0 per API call — no usage fees, no subscription.
• ChatGPT API (GPT-4o): $2.50-15.00 per 1M tokens.
• Claude API (Sonnet): $3.00-15.00 per 1M tokens.

For high-volume applications, on-device AI can save thousands of dollars per month. The trade-off is slightly lower model capability — but for many tasks (summarization, classification, translation), Gemini Nano provides sufficient quality at zero marginal cost. The total cost is just the one-time model download (~2GB storage), making it ideal for startups, internal tools, and privacy-focused products.