Ultralytics v8.4.25 is here 
Summary: Ultralytics v8.4.25 restores TensorFlow.js export for most supported environments
while also improving training efficiency, box alignment accuracy, visualization behavior, CI reliability, and documentation for Ultralytics YOLO and the Ultralytics Platform.
If you rely on browser deployment, train large pose/detection/OBB workloads, or want smoother migration guidance for YOLO26, this is a great update to pull in.
Highlights
TensorFlow.js export restored
The biggest change in v8.4.25 is the return of TF.js export for supported systems through PR #23985 by @glenn-jocher.
This update:
- Adds a
ydf<0.13.0dependency pin on non-ARM64 systems - Resolves a TensorFlow/protobuf conflict affecting export workflows
- Removes the previous hard disable on TF.js export
- Keeps ARM64 Linux blocked where support is still not ready
- Cleans up platform checks so TF.js and CoreML are handled separately
For teams deploying models to the web, this should make export workflows much more dependable again ![]()
Faster training internals
Two training-path optimizations reduce Python overhead in heavy loss computation code:
- PR #23937 by @ahmet-f-gumustas: vectorized keypoint batching for pose loss
- PR #23966 by @Laughing-q: vectorized preprocessing for
detectandobbloss calculation
These changes preserve behavior while improving efficiency, especially in larger training runs.
More accurate box alignment
PR #23967 by @a17sol fixes padding math in scale_boxes to better match LetterBox rounding.
This addresses subtle pixel-level box shifts on certain image shapes and should improve visual consistency and downstream coordinate reliability.
Better visualization control
PR #23951 by @K-saif updates visualization so show_conf=True works independently of show_labels.
That means you can now render confidence-only overlays without showing class names, which is handy for demos, debugging, and cleaner qualitative reviews.
Better docs for YOLO26 and Platform
Documentation also received a strong refresh in this release:
- PR #23947 by @raimbekovm: added an end-to-end detection guide for YOLO26
- PR #23931 by @raimbekovm: fixed
Referencesrendering in API docs - PR #23974 by @glenn-jocher: updated runner Docker images and docs-related infra
- PR #23975 by @glenn-jocher: refreshed Platform docs, including RTX PRO 4500 updates
These improvements make it easier to onboard, migrate from older YOLO generations like YOLO11, and deploy with Platform.
Improvements
CI and infrastructure updates
We also shipped a few under-the-hood improvements to keep development and testing smoother:
- PR #23875 by @Laughing-q: moved more workflows to the new
cpu-latestrunner for faster and more consistent CI - PR #23974 by @glenn-jocher: updated
Dockerfile-runnerto2.333.0
Prediction telemetry enhancement
PR #23948 by @ambitious-octopus adds backend metadata to prediction events, improving analytics and debugging visibility.
Full list of changes
- Added autobacked events name in PR #23948 by @ambitious-octopus
- Used new
cpu-latestrunner for faster CI tests in PR #23875 by @Laughing-q - Vectorized keypoint batching loop for Pose loss calculation in PR #23937 by @ahmet-f-gumustas
- Vectorized preprocess for
detect/obbloss calculation in PR #23966 by @Laughing-q - Fixed
Referencesdocstring sections rendering incorrectly in API docs in PR #23931 by @raimbekovm - Added end-to-end detection guide for YOLO26 in PR #23947 by @raimbekovm
- Updated
Dockerfile-runnertov2.333.0in PR #23974 by @glenn-jocher - Updated Platform docs with RTX PRO 4500 in PR #23975 by @glenn-jocher
- Fixed pad calculation in
scale_boxesto matchLetterBoxrounding in PR #23967 by @a17sol - Allowed
show_confto be used independently ofshow_labelsin visualization in PR #23951 by @K-saif - Added the
ultralytics 8.4.25TensorFlow.js fix withydf<0.13.0in PR #23985 by @glenn-jocher
New contributor 
A warm welcome to @a17sol, who made their first contribution with the scale_boxes alignment fix in PR #23967.
Try it out
Upgrade with:
pip install -U ultralytics
Then explore the latest release in the v8.4.25 release page, or browse the complete diff in the full changelog.
If you test this release, especially TF.js export, YOLO26 training, or Platform workflows, we’d love to hear how it goes. Thanks to everyone in the community who contributed fixes, reviews, docs, and feedback ![]()