Glance AISetup guide
Set up your local AI
Glance AI doesn't come with an AI inside it. It's a client for a model that runs on your own computer — which is exactly why your screen never leaves your PC. You supply the model once, and the app connects to it from then on.
The easiest way to get one is Ollama, a free model server for Windows. Install it, download one model, open Glance AI. That's the whole job.
ollama pull gemma3:4b in a terminal, open Glance AI. If you'd rather understand what you're choosing, keep reading — it takes two more minutes.
The three steps
-
Install Ollama
Download it from ollama.com/download and run the installer. It starts serving in the background straight away and leaves an icon in your system tray — there's no window to keep open, no
ollama serveto run, and nothing to sign up for. -
Download a model
Open a terminal — press
Win, typeterminal, pressEnter— and run:ollama pull gemma3:4bThat's about a 3.3 GB download. It's a model that can read images as well as text — which is what lets Glance AI make sense of a screenshot — and it runs on ordinary hardware, including PCs with no dedicated graphics card. Got a gaming GPU? Pick a bigger model instead.
-
Open Glance AI
It finds the server on its own and the status chip turns Ready. If the app was already open while you were downloading, click Check again.
Which model fits my PC?
gemma3:4b is the default recommendation because it runs well nearly everywhere. If you have a gaming-class graphics card, a larger model gives noticeably better answers and reads dense screenshots more reliably. Find your row:
| Your PC | Model to pull | Why |
|---|---|---|
| No dedicated graphics card, or 4 GB of video memory most laptops and office PCs |
gemma3:4bStart here |
3.3 GB. On a 4 GB card part of it runs on the processor — usable, just slower. With no graphics card at all, 16 GB of system RAM is the comfortable amount. |
| 6–8 GB of video memory RTX 4060 / 3060 laptop class |
gemma3:4b |
Fits your card completely — fast and snappy. The 12B model wouldn't, and would slow to a crawl. |
| 10–16 GB of video memory RTX 3060 desktop / 4070 / 5070 class |
gemma3:12b |
8.1 GB. Clearly better reasoning, and more reliable at reading busy or low-contrast screenshots. |
| 24 GB or more of video memory RTX 4090 / 5090 class |
gemma3:27b |
17 GB. The best quality you can get without sending anything off your machine. |
Those figures are download sizes. Running a model needs a bit more room than that on top, so leave some headroom rather than picking the largest model that technically fits.
Not sure how much video memory you have? Press Ctrl+Shift+Esc for Task Manager, open Performance, and click your GPU — "Dedicated GPU memory" is the number you want. If you don't see a GPU section beyond the built-in one, take the first row.
There's nothing smaller than gemma3:4b for this. Gemma 3 does come in 1B and smaller sizes, but those are text-only — they can't look at a screenshot, which is the entire point of Glance AI. 4B is the floor.
Need your GPU for other things too? Games, video editing, and streaming compete for that same video memory. If you use them while Glance AI is open, go one row up — gemma3:4b is a perfectly good everyday choice even on a high-end card, and it leaves the rest of the GPU free.
Switching later is easy. Pull the new model in a terminal, then open Settings → Model in Glance AI — every model your server has appears in the list. Pick one and you're done. If answers ever start feeling slow, the model is probably too big for your card: drop down a row.
Good to know
- Voice input needs a different kind of model. The Gemma 3 models handle text and images, not audio, so the mic button won't work with them — Glance AI tells you in the chat rather than failing silently. Typed questions and screen captures always work.
- Already running LM Studio, llama.cpp, or Jan? Then you don't need Ollama at all — Glance AI scans the usual local ports and connects to whichever server it finds (Ollama
11434, LM Studio1234, Jan1337, llama.cpp8080). Two catches: unlike Ollama, those three don't start on their own — you have to start the server in the app each time, and nothing can find a server that isn't running. And whichever you use, the model you've loaded there has to accept images. - No PC that can handle it? You can point Settings → Endpoint at any OpenAI-compatible server instead. Be aware of the tradeoff: that sends your questions and screen captures off your device to whoever runs it. The app marks the endpoint as online when you do, so the state is never ambiguous.
- Nothing here reaches us. BroMind runs no servers and receives nothing either way — see the privacy policy.
Troubleshooting
The model server isn't reachable. Look for the Ollama icon in your system tray — if it isn't there, launch Ollama from the Start menu. To confirm it's really answering, run ollama list in a terminal: it prints your installed models if all is well, and an error if nothing is listening. Then click Check again in Glance AI.
Open Settings → Model, pick it from the list, and hit Test connection. The app doesn't switch models on its own.
Usually the model doesn't fit entirely in your video memory, so part of it is running on the processor. Switch to a smaller model — one row up in the table above. If you have no graphics card at all, expect a pause of several seconds before an answer starts, especially when you attach a screenshot: a picture is a lot of input to chew through, and that part is slowest on a processor.
The model you've selected is text-only. Vision is a per-model capability, not something the app can add. Pull one of the Gemma 3 models above, which all read images.
Write to contact@bromind.dev — a human reads it.