Most AI models need at least 24 but preferably 32.
Where are you getting this information from? Most models that are less than 16B params will run just fine with less than 24 GB of VRAM. This github discussion thread for open-webui (a frontend for Ollama) has a decent reference for VRAM requirements.
Can you elaborate here? I've had very few issues with Flatpaks and the documentation is pretty thorough. I'm curious what wider issues it has to make the whole ecosystem "pretty shit" and unreliable.