New Release: Ultralytics v8.3.111

Discussion
,
1

:rocket: Ultralytics v8.3.111 Release: Faster YOLOv10, Slicker Tutorials, and Smarter Integrations!

:glowing_star: Summary

We’re excited to announce Ultralytics v8.3.111—a release full of efficiency boosts, smoother onboarding, and improved integrations. This update supercharges YOLOv10 models for speed and size, enhances user experience in our tutorials, and brings even more reliability to your workflows and exports. Read on for all the details!

:new_button: New Features & Improvements

:high_voltage: YOLOv10 Gets Even Faster and Lighter

Our core focus this release is optimizing YOLOv10 for real-world deployment:

  • Automatic Detection Head Removal: The fuse() method now intelligently drops the “one2many” detection head when not required. This means:
    • YOLOv10n now trims down from 125 to just 102 layers
    • Parameter count shrinks from 2.76M to 2.30M
    • Instant speed and efficiency gains, especially on edge and production devices

Pull Request: YOLOv10 skip one2many head when fused by @Y-T-G

:construction: Improved Export and Integration

  • TensorRT Export for YOLO-World: YOLO-World models can now be exported to TensorRT with more robust dynamic shape handling, making deployment on NVIDIA hardware easier and more reliable.

  • TensorBoard Logging Toggled Off by Default: For users who prioritize fast startup, TensorBoard is now off by default, easily enabled when you need detailed experiment tracking.

  • Cleaner Notebook Output: Say goodbye to distracting Albumentations warnings in your training notebooks for a smoother Colab/Jupyter experience.

:books: Slicker Tutorials & Stronger Docs

  • Revamped YOLO11 Colab Notebook: We’ve updated instructions, visuals, and links—making it easier than ever to get started, and ensuring everything is up-to-date with our rapidly growing documentation and community resources.

  • Automated Link Checking: Documentation builds now fail on broken links, guaranteeing you always have up-to-date working resources.

:bullseye: Why This Matters

  • Speed and Efficiency: Lighter YOLOv10 models mean less memory consumption and faster inference, which is especially valuable for production and edge devices.
  • Flexibility in Experiment Tracking: Choose the logging tool that best fits your workflow, with even easier setup.
  • Better Onboarding: Our improved tutorials and documentation make getting started and staying productive seamless for everyone.
  • Quality and Reliability: Automated doc checking and export fixes keep your pipeline and resources robust.

:busts_in_silhouette: Contributors

A special welcome and thanks to our new contributor:

  • @laugh12321 made their first contribution with improved TensorRT export support!

And thanks to @Y-T-G and @glenn-jocher for ongoing improvements.

:link: Details & How to Get Started

Full changelog: v8.3.110…v8.3.111 comparison
Release page: Ultralytics v8.3.111 release notes

:raising_hands: Try It Out and Share Feedback!

Upgrade to v8.3.111 and let us know what you think—your feedback is crucial for the YOLO community and the continued success of Ultralytics! Whether you’re deploying at scale or just experimenting, we’re eager to hear how these updates help you.

If you have questions or want to share your results, just reply to this post or open a GitHub Discussion.

Thank you for being part of our open-source journey and for helping advance state-of-the-art vision AI!