architectds / collabosm

# collabosm — user guide
collabosm lets you use a large AI model — **Qwen3.8-27B**, or **Qwen3.8-Flash-Next** — running on
a powerful GPU that is rented in **your own Google Colab account**. You chat with it in a simple
app on your computer, start and stop the GPU with a button, and see the price before you pay.
Other programs, such as Codex or ModelDock, can use the same model.
You pay Google for the GPU only while it runs. collabosm itself is free and open source.
---
## Quick start
1. Make sure your Google account has **Colab compute units** (a paid plan, or pay-as-you-go).
2. Install **Python 3.12 or newer**.
3. Download collabosm, open a terminal in its folder, and run `python frontend/collabosm.py`. The page
 (**http://127.0.0.1:3020**) opens in your browser by itself, and a **collabosm** shortcut appears
 on your desktop for next time.
4. Follow the **Colab** steps on the right: **Install**, then **Connect Google account**.
5. Under **GPUs** press **Start**, then **Start — billing begins**. A few minutes later (about 6 for
 the default A100-40G recipe, 11 for the A100-80G) the panel says **Ready**: chat. The A100-40G
 Strata recipe is the exception: allow up to its 40-minute estimate. The experimental 80G
 Maya-L option has a 110-minute first-run estimate. Press **Stop** when you are done.
Every step is explained below.
## Contents
1. Before you begin
2. Install
3. Connect your Colab (first time only)
4. Everyday use: open the app, start a GPU, chat, **stop the GPU**, switch
 models, see what you spent, check the connection
5. Use the model in other programs: Codex, ModelDock,
 Open WebUI, your own scripts, another device
6. Troubleshooting
7. Update or uninstall
- Appendix: try it without paying, what it costs, the command line, how it works and
 your privacy, for developers and AI agents
---
## 1. Before you begin
### A Google account with Colab compute units
The model needs an **A100** GPU, and Colab rents those only for **compute units**: a paid Colab plan
(Pro or Pro+) or units bought pay-as-you-go. The free tier is not enough. Get them at
colab.research.google.com/signup.
How much? The GPU this uses costs **6.77 compute units per hour, about $0.68**. 100 compute units
run it for about 14 hours. See what it costs.
### Python 3.12 or newer
- **Windows:** download Python from python.org. In the installer,
 tick **"Add python.exe to PATH"** before you press Install.
- **macOS:** download Python from python.org — the `python3`
 that comes with macOS is too old — or run `brew install python@3.12` if you use Homebrew.
- **Linux:** install it with your package manager, or use uv.
To check: open a terminal (below), type `python --version` and press Enter. It should say 3.12 or
higher.
### Opening a terminal
A terminal is a window where you type commands.
- **Windows:** open the Start menu, type **PowerShell**, press Enter.
- **macOS:** press ⌘+Space, type **Terminal**, press Enter.
- **Linux:** usually Ctrl+Alt+T.
In this guide, a command suc