Ultralytics v8.4.16 is out! 
TL;DR: Ultralytics v8.4.16 significantly improves Windows ↔ Linux/macOS .pt model portability by fixing pathlib pickle/unpickle edge cases, plus a small docs refresh around HEIC/HEIF support.
New Features / Highlights
Cross-platform .pt loading fix (Windows ↔ Linux/macOS)
If you’ve ever hit errors like “can’t unpickle WindowsPath/PosixPath” when moving weights between machines/CI, this release targets that directly.
- PR #23725 — Windows<>Linux model compatibility fix by @glenn-jocher
torch_safe_load()now remapspathlib.WindowsPath↔pathlib.PosixPathduring unpickling- Prevents common
.ptload failures when models are trained on one OS and used on another
Improvements
Safer, more consistent checkpoint metadata
model.pt_pathis now stored as a plain string (str(weight)) rather than aPathobject- This improves consistency across OS boundaries and helps downstream tools that rely on
model.pt_path
Docs Updates
HEIC/HEIF + AVIF format clarity
- PR #23717 — Update HEIC/HEIF dependency info and add .heif format to docs by @raimbekovm
- Docs now clarify HEIC/HEIF uses
pi-heif(auto-installed on first use) - AVIF is supported natively by Pillow
.heifis now explicitly listed among supported image formats
- Docs now clarify HEIC/HEIF uses
Version Bump
8.4.15→8.4.16
Why this matters
- Fewer cross-OS
.ptloading headaches when sharing weights with teammates
- Smoother reproducibility in mixed environments (Linux training → Windows testing/deploy, or vice versa)
- Clearer HEIC/HEIF dependency expectations for a more seamless image pipeline
Update now
pip install -U ultralytics
Try it out and let us know what you see—especially if you’ve had cross-platform .pt portability issues before. Your feedback helps keep Ultralytics YOLO and the wider YOLO community moving fast and staying reliable. 
You can review the release details in the v8.4.16 GitHub release and browse the full v8.4.15…v8.4.16 changelog.