What it is
modelPlayground is our kit for running open AI models on a home computer with the network off, and switching to cloud models when we choose. One command starts a chat window in the browser, with offline search over your own documents, and another command shuts it down to free memory. Coding helper scripts work fully local, fully cloud, or hybrid, where a cloud model plans and a local model makes the edits. A "rainy day" guide covers what to do when cloud AI services are down. It is personal infrastructure, not a product anyone else can install yet.
How it works
- Run the start command; the chat window opens on the local machine.
- Pick a local model and chat, or ask questions about documents in the offline library.
- For coding, launch the coding helper in local, cloud or hybrid mode.
- Run the stop command to free memory when done.
What exists today
- A start, status, stop and GPU script for the whole stack
- Scripts to download models, import them and set up the graphics chip
- Coding helper launch scripts for the three modes
- An offline document-search setup and a large offline reference library with its own server script
- Model surveys from September and October 2026 comparing open models
- The rainy-day guide and a stack overview
What isn't there yet
- Not packaged; scripts assume one specific machine
- The newest model lineup needs an upgrade of the model runner first
- No screenshots or walkthrough
- Models and the library take hundreds of gigabytes, so it can't be demoed on the web
Under the hood
- Ollama runs the models; Open WebUI (in a container) is the chat window
- Aider is the terminal coding helper
- A llama.cpp build uses Vulkan to run models on a built-in laptop-class graphics chip
- Cloud access goes through OpenRouter, one service that reaches many models