Chrome 148 Can Now Run an AI Model Inside Your Tabs — No API Key, No Cloud
Everyone assumes browser AI needs a server and an API key. Chrome 148 quietly ships Gemini Nano directly inside the browser, so inference runs on your own machine — offline, private, and fast.
01. What It Is
Chrome 148, released May 5, 2026, exposes a new Prompt API that lets web pages and extensions talk to Gemini Nano — Google's lightweight on-device language model — without ever making a network request. Instead of sending text to a cloud endpoint, the model runs directly on your computer's hardware.
Because there's no server call, there's no API key to manage, no per-request billing, and no data leaving your device. The first time an origin uses the API, Chrome downloads the model (roughly 2–4 GB) once, and after that inference happens locally — even offline. It's optimized for focused, single-turn tasks like summarizing, rewriting, and translating short passages.
This flips a common assumption on its head: most people believe every 'AI feature' on a website is powered by a remote server. Chrome's Prompt API proves the browser itself can be the AI runtime, which changes the math on cost, latency, and privacy for a whole class of small tasks.
Why It Matters
For everyday small tasks — cleaning up a note, summarizing a paragraph, quick rewrites — you get instant, private results with zero cloud cost. Developers save money by skipping API bills entirely, and users keep their text on their own machine. Because it runs offline, it also works on a plane or with a spotty connection.
Who Can Benefit
- Web developers who want cheap, private AI for lightweight features without paying per token
- Privacy-conscious users who don't want their text sent to cloud servers
- Extension builders who need on-device inference that keeps working offline
- Tinkerers who want to experiment with a local LLM straight from the browser console
02. Step-by-Step Guide
- 1
Confirm you have official Chrome 148+ on a supported desktop
This only works in the official Google Chrome build, version 148 or later, on Windows 10+, macOS 13+, or Linux. It does NOT work on Android, iOS, ChromeOS, or distro-packaged Chromium builds. Check your version at chrome://version.
- 2
Make sure you meet the hardware and storage requirements
You need at least 10 GB of free disk space (the model auto-deletes if free space later drops below 10 GB). A GPU with 4+ GB VRAM is recommended for fast inference; without it, Chrome falls back to CPU, which works but is slower.
- 3
Enable the on-device model flag
Go to chrome://flags/#optimization-guide-on-device-model and set it to 'Enabled'. This tells Chrome it's allowed to fetch and run the local model.
- 4
Enable the Prompt API flag
Go to chrome://flags/#prompt-api-for-gemini-nano and set it to 'Enabled'. Then click 'Relaunch' to restart Chrome with both flags active.
- 5
Trigger the one-time model download
Open DevTools (F12) → Console and call the API for the first time (for example, create a language model session). Chrome downloads Gemini Nano in the background — this is a one-time 2–4 GB fetch. After it finishes, all future calls run locally.
- 6
Run a focused prompt locally
Use the API for short, single-turn tasks like summarizing or rewriting a paragraph. Keep your input under roughly 2,000 tokens for good results — this is a small model tuned for narrow tasks, not deep reasoning.
Pro Tips
- Keep prompts short and single-purpose: summarize this, rewrite this more formally, translate this sentence. The model shines on focused tasks and struggles with long or open-ended ones.
- For a real website (not just local testing), flags won't help visitors — register an Origin Trial token for your domain, or ship your feature as a Chrome Extension, which can also use the API inside its ServiceWorker.
- As of Chrome 149 the model handles English, Spanish, Japanese, German, and French, so you can build lightweight multilingual rewriting and translation features on-device.
- Since everything runs locally, it keeps working offline — great for note-cleanup or draft rewriting on a flight.
Warnings & Limitations
- Gemini Nano is a small model: don't rely on it for factual accuracy, deep reasoning, long-context analysis, or code generation. Use cloud models for those.
- The context window is only about 2,000 input tokens — exceeding it noticeably degrades quality, so chunk long text yourself.
- Desktop only. There is no support on Android, iOS, or ChromeOS, and it requires the official Chrome build, not packaged Chromium.
- In Chrome 148 the API is not available in Web Workers, Shared Workers, or page ServiceWorkers (though it does work in extension ServiceWorkers).
- If your free disk space drops below 10 GB after download, Chrome deletes the model and you'll have to re-download it once space is available.
Related Tricks
Most Android Users Don't Know Gemini Can Build Custom Widgets From a Sentence
Describe the widget you want in plain English, and Gemini designs, builds, and installs it on your home screen—no settings, no widget menu digging.
Most Claude Users Miss This Ghost Icon That Turns Off Memory Instantly
Claude quietly added an Incognito mode that skips memory entirely—no chat history, no memory summaries, no training. It's a one-click privacy escape hatch hiding in plain sight.
Most Claude Users Don't Know It Can Query Your Live Database
Stop exporting CSVs and pasting query results. Claude can now connect directly to your PostgreSQL or MySQL database and run read-only queries in real time—analyzing production data without ever leaving the chat.