Type a question in Chrome’s new tab and a Qwen3.5 model answers from inside your browser, on your GPU through WebGPU or on your CPU. It downloads once and then it is a file on your disk. Your questions go nowhere. Turn off the Wi-Fi and it keeps answering.
Free. Open source. One model download, 533 MB to 2.7 GB, cached for good.
Every step, and where it happens. In local mode none of them involves a server, because there is none.
| You do | Fogar does | It happens |
|---|---|---|
| Open a tab | Nothing, yet. A new tab does not load the model on its own.Opening tabs to type an address stays free. A tab nobody has looked at for five minutes gives the memory back. | This tab |
| Ask | Wakes the cached model from your browser’s private file storage and streams the answer.Arithmetic is computed, not guessed. Addresses open. “remind me…” and “todo:” go where they belong. | This device |
| Follow up | Sends the conversation so far along with the new question.“Shorter”, “now as an email”, “and if I already pushed it?” all work. The oldest turns drop off to fit the 4,096-token window. | This device |
| Read | Renders the answer as Markdown: lists, bold, code, links. One click copies it.The system prompt tells the model to say when it does not know, and never to invent names, dates, quotes, or numbers. | This tab |
| Check | Counts every request the page made, by host, in Settings under Network, this tab.Read from the browser’s own timeline. Nothing inferred. | You can see it |
No account. No analytics. The conversation lives in the tab and leaves with it.
The first screen looks at your GPU and memory and recommends a tier. You can pick another any time in Settings.
recommended, never forcedQwen3.5, quantized, from Hugging Face. It lands in your browser’s private file storage (OPFS) and never downloads again.
533 MB · 1.3 GB · 2.7 GBThe model runs through wllama, llama.cpp compiled to WebAssembly, on your GPU through WebGPU or across your CPU cores. If WebGPU fails, it falls back to the CPU on its own.
warm load 1.3 to 3.2 s · 44 to 92 tok/s · Apple silicon, WebGPUQuick rewrites and summaries. Loads in about a second once cached. The pick for machines without WebGPU.
92 tok/sNoticeably smarter. Recommended on WebGPU machines with 8 GB of memory or more.
71 tok/sThe first tier that behaves like an assistant: it got the fact, the arithmetic, and the sorting right in our tests. Recommended on Apple silicon.
44 tok/sTick one box and Fogar looks the question up before answering, then cites what it found. Free with no key, or full web results with your own key. Answers with sources.
Point Fogar at any OpenAI-compatible endpoint: OpenAI, OpenRouter, Groq, or Ollama on localhost. The key stays in this browser and goes only to the address you typed.
Drop a PDF, Word, or text file on the page. It is read in the tab, never uploaded, and the answer comes from it. About attachments.
Rewrite, reply, summarize, explain. Forms for each, and ones you make yourself. About recipes.
For everyday questions, rewriting, and explaining, yes: the model runs in your browser and nothing you type is sent anywhere. It is much smaller than ChatGPT, so it knows less. When you need more, switch to Cloud with your own key, or tick “Search the web first” for facts.
Yes, in local mode, once the model is cached. Turn off Wi-Fi and it keeps answering. Web search and cloud mode need a connection, by definition.
No. There is no Fogar account and no server. A key is needed only if you choose to use a cloud model or a paid search provider, and it stays in this browser.
Qwen3.5 in three sizes: 0.8B (533 MB), 2B (1.3 GB), and 4B (2.7 GB), quantized to 4 bits. Fogar recommends one from your GPU and memory. A Qwen3 0.6B is there for machines with very little memory.
The model loads only when you ask something, and a tab nobody has looked at for five minutes releases it. Loading from disk takes one to three seconds. Without WebGPU it runs on the CPU, slower but working.
Open Settings and look at Network, this tab. It counts every request the page made, by host, from the browser’s own timeline. The code is open source too, so you can read what the page does.
Fogar is on the Chrome Web Store. One click to add it, and it updates itself from there. Then open a new tab.
Free, from the Chrome Web Store. Brave, Arc, and Edge install from it too.
On your phone, or not on Chrome? The same Fogar runs as a web app at app.fogar.ai. Open it once with a connection, add it to your home screen (Safari: Share, then Add to Home Screen; Chrome: the menu, then Add to Home screen), and set it up from there; after that it works offline. The widgets that need Chrome’s own APIs, Sessions, bookmarks, Inbox, Feed, Jira, and Agenda, stay on the desktop.
Loaded the zip before the listing went live? The store copy has a different extension ID, so your data will not follow on its own. Three steps move it over.
Todos, notes, recipes, widgets, sessions, and settings go into one file.
It opens on your next new tab, empty.
Once it looks right, remove the old copy in chrome://extensions.
Can't use the store? The same build is here as a zip: unzip it, turn on Developer mode in chrome://extensions, and Load unpacked. It will not update itself.