I replaced the old system of 2 models with a new system where the desktop assistant only uses a single Gemini Live model. This results in faster response times and higher API call limits. You can find it at https://github.com/smitamw/esp32s3-assistant. (The image is unrelated.)