You were definitely hitting the limitations of the coral. Fregata (a native port of Frigate to macOS) can easily handle 25+ cameras since Apple Silicon can do detections in 1-2ms, giving you 500-1000 object detection inferences per second.
Wireless cameras can also cause their own set of issues, but I can understand using them if you have to.
I've been a fan and avid user of the excellent privacy-first local Frigate NVR for years (https://frigate.video if you're unfamiliar). For the last year or so I've been running Frigate as a docker container on my Mac Mini. While this setup worked, and was kind of energy efficient, CPU usage was high and it felt like I was leaving a LOT of the performance and benefits of Apple Silicon on the table.
As many of you know, Docker on Mac runs containers inside of a Linux VM, blocking container access to the Apple Neural Engine (ANE), GPU, and the hardware media decoding/encoding Media Engine. As well as adding VM overhead, this prevents Frigate running on docker from using any of that hardware for things like AI object detection, hardware accelerated video stream decoding/encoding, Frigate's AI Enrichments, and more.
Enter Fregata (Fregata is the scientific genus that frigatebirds belong to).
Fregata ports the runtime Frigate needs (think process supervisor, RAM disk management, environment variables, inter-process communication, shared memory, etc) that is normally provided by Docker into a native Mac .app, signed and notarized by Apple.
Additionally we added and maintain a Apple Core ML AI object detector written in Rust that uses the ANE and hardware-accelerated frame cropping, plus hundreds of other Frigate optimizations and tweaks specific to macOS.
Fregata is the runtime that RUNS a macOS-customized version of Frigate. We'll keep Fregata up to date with upstream Frigate releases.
All Frigate features work as usual (including Frigate+), just accelerated by Apple Silicon.
The result is AI object detection inference times of 1-2ms on the ANE, stream decode/encode accelerated by Apple's Media Engine, GPU-accelerated Frigate enrichments, very minimal CPU usage, and the extreme energy efficiency of the Mac Mini (or any Apple Silicon Mac you have laying around).
We just published our first GA release. Any and all user feedback is very welcome and happy to answer any questions you have. If you're a Mac user that runs Frigate, we think you'll be very pleased with Fregata.
The Qwen models seem to be the best currently in my experience.
But if you're thinking about buying new hardware just for local inference: don't.
You'll spend more upfront for rapidly out of date and depreciating hardware while running inferior models compared to the latest from Anthropic and OpenAI.
Unless your requirement is prompt privacy, or you have idle hardware sitting around, the ROI just isn't there compared to a monthly subscription.
Surprisingly I'm seeing a rapid increase of discoverability for my apps FROM LLMs. The flood of Claude/ChatGPT users finding my apps right after I get a basic site up, before I've told anyone it exists, is astonishing.
If you are solving a problem users are searching for, you will be found and found faster than ever via LLMs.
Wireless cameras can also cause their own set of issues, but I can understand using them if you have to.