Run a private AI chatbot in a browser tab, with one click
Click a link, wait a minute, and a real AI model is running on your laptop's graphics chip inside Chrome or Edge. Nothing to install, no account, and once it has loaded you can switch the Wi-Fi off and keep chatting.
- Real speed on an ordinary laptop with no gaming graphics card: 31 to 40 tokens a second
- Chrome or Edge, version 113 or newer, on Windows or Mac (other browsers and phones: see step 1)
- 8GB of RAM minimum, 16GB recommended. With the smallest model loaded, the browser uses about 1.8GB extra.
- About 1GB of free disk space for a small model (bigger ones need more). It's stored inside the browser, not as a normal file.
- Wi-Fi rather than mobile data: the first load downloads about 700MB
Open your private AI page and ask it something
One click and about a minute of downloading, and you're chatting with an AI that runs on your own graphics chip.
Open [Private AI in a tab](/learn/ai-in-a-browser-tab/private-ai-tab.html) in Chrome or Edge. It's a single web page built on WebLLM, an open-source engine that runs AI models with WebGPU: a browser feature that lets a page do heavy maths on your graphics chip. The Intel graphics built into an ordinary laptop is enough.
The top box tells you straight away whether your browser can do it. Then:
- Leave Llama 3.2 1B selected and click Load model.
- Watch the progress bar. The first time, it downloads about 700MB of model files from Hugging Face (the big public library of AI models). On home broadband that takes about a minute.
- When it says Ready, click Ask.
First run: 42.5 seconds including the 705MB download, then 32 tokens a second. Opened again later, the same model was ready in 2 seconds. The grey line under the answer is the speed. On a normal laptop with Intel graphics this page writes 31 to 40 tokens a second. A token is roughly three quarters of a word, so that's faster than you can read.
Which browsers have WebGPU (from the WebGPU status page ↗):
- Chrome and Edge: on by default on Windows and Mac since version 113. This is the path this mission uses.
- Firefox: on by default on Windows since version 141, and on Macs from version 145 or 147 depending on the Mac.
- Safari: on by default from macOS 26 and iOS 26.
- Linux: Chrome has it on by default for Intel Gen12 graphics and newer (most Intel laptops from 11th-generation Core onwards) from version 144, and for NVIDIA on Wayland from version 147 with driver 535.183.01 or newer. Other graphics chips need launch flags, listed on the status page. Firefox on Linux only has it in Nightly.
- Phones: Chrome on most Android 12 or newer phones, and Safari on iOS 26. Open the same link on your phone, pick SmolLM2 360M first (about 0.2GB) and step up to Llama 3.2 1B if it copes. Phones have far less memory to spare than laptops, so close other apps first.
If it breaksThe page says "No WebGPU in this browser" + 3 more
The page says "No WebGPU in this browser"
The download stops, or the tab crashes halfway
It's very slow, or the whole computer stutters
The top box says it has no f16 support
q4f32_1 model there. It's slower, but works on more hardware.