New Release: Ultralytics v8.4.75

Discussion
, ,
1

:rocket: Ultralytics v8.4.75 Released

Quick summary: Ultralytics v8.4.75 is a small but high-impact release focused on CoreML reliability and speed on macOS :red_apple::high_voltage:. CoreML models now use the CPU + Neural Engine path by default on supported Macs, avoiding a crash-prone setting and delivering a much smoother local inference experience for Ultralytics YOLO exports such as YOLO26 and YOLO11.

:glowing_star: Highlights

  • :ambulance: Major CoreML stability fix for macOS
  • :high_voltage: Apple Neural Engine now used by default on supported Macs
  • :shield: Automatic fallback for older macOS versions
  • :memo: Updated docs explaining the new CoreML behavior
  • :white_check_mark: No new model architecture changes — this release is all about backend reliability and performance

New Features

:red_apple: CoreML now defaults to CPU + Neural Engine on macOS

In v8.4.75, the CoreML backend now loads models with ComputeUnit.CPU_AND_NE instead of the previous default behavior.

This change is especially important because certain CoreML execution paths on macOS could trigger hard Python crashes with errors like MLIR pass manager failed. By switching to the CPU + Neural Engine path, Ultralytics improves both stability and performance for local Mac inference.

:high_voltage: Faster inference on supported Apple hardware

On macOS 13 and newer, CoreML inference now automatically uses the Neural Engine when available. Reported results show roughly:

  • 2.5 ms on Neural Engine
  • 8.5 ms on CPU

That’s about a 3x speedup, which is a great improvement for real-time and interactive applications.

Improvements

:shield: Better compatibility on older macOS versions

If CPU_AND_NE is not supported on the host system, Ultralytics now falls back to CPU_ONLY instead of failing. This makes CoreML deployment more dependable across a wider range of Mac environments.

:memo: Clearer CoreML documentation

The CoreML integration docs were updated to explain the new macOS behavior and why avoiding the GPU path currently improves stability. This should make troubleshooting and deployment much easier for Apple Silicon users.

Bug Fixes

:wrench: Fixes crash-prone CoreML execution paths on Mac

Before this release, running a CoreML .mlpackage from Python on macOS could crash outright in some environments. With v8.4.75, that workflow should now work out of the box on supported Macs with a much more reliable default configuration.

Pull Requests

Impact

This release will be especially useful if you are:

While this is not a new model release, it’s an important quality-of-life update for Mac users and should make CoreML workflows much more dependable :rocket:

Try It Out

Update with:

pip install -U ultralytics

Then test your existing CoreML export workflow and let us know how it performs on your Mac.

You can explore the release in the Ultralytics v8.4.75 release page, and review everything included in the full changelog for v8.4.75.

Feedback

Please give v8.4.75 a try and share your results, especially if you’re running CoreML inference on Apple Silicon. Feedback from the community helps the Ultralytics team keep improving the experience for everyone :raising_hands: