Ultralytics v8.4.150 Released
Ultralytics v8.4.150 delivers faster and more secure checkpoint loading, restores fused YOLOE checkpoint compatibility, improves RT-DETR training efficiency on dense datasets, and expands model and deployment documentation.
Highlights
Faster restricted checkpoint loading and reduced cold-start overhead
Restored loading and warmup support for fused YOLOE checkpoints
More efficient RT-DETR matching for images containing many objects
Improved privacy for Sentry diagnostics
Expanded Ultralytics Platform deployment guidance
A clearly labeled YOLO27 preview, while YOLO26 remains the recommended released model
Clearer benchmark tables across supported model families
New Features and Documentation
YOLO27 Preview Documentation
PR #26154 by @glenn-jocher introduces a YOLO27 Coming Soon page covering planned architecture details, supported tasks, preliminary benchmarks, and future usage examples.
YOLO27 weights, package support, and a launch date are not currently available. For new projects, YOLO26 remains the recommended stable Ultralytics model.
Expanded Ultralytics Platform Deployment Guidance
PR #26149 by @sergiuwaxmann documents:
- Configurable CPU and memory resources
- Uptime billing and scale-to-zero behavior
- Lightweight monitoring for eligible paid endpoints
- Temporary prediction examples and statistics
- Saving useful prediction examples to datasets for persistent storage
The Ultralytics Platform provides an integrated workflow for annotating datasets, training models, and deploying Ultralytics YOLO.
Performance Improvements
Faster Restricted Checkpoint Loading
PR #26155 by @glenn-jocher resolves only the model classes referenced by a checkpoint instead of scanning every supported model module.
This update:
- Reduces unnecessary module discovery and startup overhead
- Preserves trusted namespaces, aliases, registration locks, and allow-list behavior
- Restores safe
forwardandforward_fusebindings for fused YOLOE checkpoints - Improves rejected-object error messages without calling
repr()on partially reconstructed modules - Expands tests for legacy checkpoints, fused models, and inference consistency
Faster RT-DETR Matching on Dense Datasets
PR #26150 by @fcakyon avoids computing matching costs between predictions and targets belonging to different images.
This reduces unnecessary storage and padding overhead, with the largest benefits expected on dense-object datasets such as SKU-110K. The assignment solver, losses, precision, and training configuration remain unchanged.
Clearer Model Benchmarks
PR #26156 by @glenn-jocher standardizes bold highlighting for leading accuracy, speed, latency, size, parameter-count, and computational-cost values.
Tables were updated across YOLO26, YOLO11, YOLOv8, YOLOv9, YOLO12, YOLOE, SAM, YOLO-NAS, and other supported families without changing the underlying benchmark results.
Compatibility and Bug Fixes
Legacy NumPy Metadata and CLIP Support
PR #26153 by @glenn-jocher adds restricted-loading support for legacy NumPy metadata and embedded Ultralytics CLIP objects.
Together with the fused-checkpoint updates, this improves compatibility with older Platform checkpoints while maintaining explicit loading restrictions.
Safer Diagnostics and Updated Ray Tune Support
PR #26151 by @glenn-jocher prevents Sentry from capturing command-line arguments and local variables that may contain credentials.
The update also requires Ray 2.41.0 or newer and removes obsolete compatibility logic for older Ray releases.
Upgrade and Try It
Upgrade to v8.4.150 with:
pip install -U ultralytics
Explore the complete release details on the Ultralytics v8.4.150 release page, or review every change in the v8.4.149 to v8.4.150 comparison.
Please try the release and share your feedback, benchmark results, or any issues you encounter. Thanks to every contributor and the wider YOLO community for helping improve Ultralytics YOLO! 