Mac mini home camera AI keeps hallucinating — model choice or settings issue?
Someone is running a 16GB M4 Mac mini as a home server, with using about 3GB of RAM and Pi-hole using 1GB, leaving roughly 8-9GB free for a local vision AI model. Nine cameras are set up around the house to trigger image analysis on motion, sending a push alert through when a person, animal, or vehicle is detected.
The setup runs into problems: some models fail outright without enough allocated tokens, while others take around 15 seconds per analysis. The glimpse-v1 model gives reasonably solid results but lacks detail.
The qwen2.5vl:3b model currently running on two test cameras gives the most sensible output overall but frequently gets details wrong — for instance, it sometimes reports a parked, stationary car as moving and parallel parking, and it often misidentifies a wooden support structure as something else. The person is looking for a better model or settings that can cut analysis time below 15 seconds while improving accuracy.
Key points
- 16GB M4 Mac mini running (~3GB) and Pi-hole (~1GB), analyzing motion-triggered images from 9 cameras
- glimpse-v1 model gives decent but not very descriptive results
- qwen2.5vl:3b is the most accurate so far but , e.g. reporting a stationary car as moving and parallel parking
- Analysis currently takes up to 15 seconds per image; looking for faster settings or a better model