@BennettBuhner i think 8b ones would be slow since macOS already uses 5-7GB of RAM idle – when i use the smaller gemma/qwen models locally i always go for 1-3b param models, 4b+ is pushing it