Switch off your Wi-Fi. The AI keeps working.
This page loads a small language model into your browser. Once it has loaded, disconnect from the internet and give it a task: it summarises, answers and extracts on your own device, and the counter shows the network requests made while it worked. It is the same idea as a private AI deployment, scaled down to fit in a browser.
Step 1: load the model into your browser
A one-time download of roughly one to two gigabytes, stored in your browser's cache. It needs a desktop browser with WebGPU (recent Chrome or Edge, or Safari 26) and a few gigabytes of free graphics memory.
Check it yourself
- Load the model, then open your browser's developer tools (F12, or Option-Cmd-I on a Mac) and choose the Network tab.
- Switch off Wi-Fi or unplug the network cable. The status above changes to "offline".
- Run a task. The answer appears while the Network tab stays empty, because no request is made.
- Paste your own text into the document box. It never leaves the page.
What this proves: a language model can do useful work with no connection at all. What it doesn't prove: production quality. A private deployment runs larger models on your own server or an in-country cloud, with access control, audit logs and an evaluation on your own documents. See how that works, what it costs and how pilots run on the private AI page.
Questions about the demo
Does anything I type leave my computer?
No. After the one-time model download, the model runs on your device's graphics processor through WebGPU. The counter shows how many network requests the page made while each answer was generated, and you can check the same thing in your browser's developer tools.
Why is the answer quality lower than a cloud chatbot?
This page runs a small model (1.7 or 4 billion parameters) so it fits in a browser. A private deployment on a server runs much larger open-weight models, typically 8 to 120 billion parameters, on your own hardware.
Which model is this?
Qwen3 from Alibaba, released under the Apache 2.0 licence, running through the open-source WebLLM engine. For client deployments we usually lead with IBM Granite and, for Arabic, Gulf-built models such as Jais, ALLaM and Fanar.
Does it work in Arabic?
Small browser-sized models are weak in Arabic, so this demo uses English tasks. Arabic-English drafting runs on the larger Arabic models in a server deployment, which we can show you in a live session.
Why does it need WebGPU?
WebGPU lets the browser use your graphics processor to run the model at a useful speed. Recent desktop Chrome and Edge support it, and so does Safari 26. Without it the page explains how to book a live demo instead.