New Release: Ultralytics v8.3.116

Discussion
,
1

:rocket: Ultralytics v8.3.116 Release Announcement

:glowing_star: Summary

We’re excited to announce Ultralytics v8.3.116 — a feature-rich release designed to make your computer vision journey smoother, more flexible, and more productive than ever! This update introduces advanced loss customization, more robust export options, improved NMS reliability, and wide-reaching usability and performance enhancements across the platform.

Read on to discover what’s new, how it benefits you, and where to find even more details.

:bar_chart: Key Changes

New Features

  • Customizable Loss Functions
    Now featuring gamma and alpha parameters for FocalLoss and VarifocalLoss, offering fine-grained control over how your models handle class imbalance and focus on challenging samples.
    See PR by @pow3rpi

  • TorchScript Half-Precision Export
    Models can now be exported in half-precision via the half argument when using TorchScript, enabling faster inference and reduced memory usage on compatible hardware.
    See PR by @seungjlee

  • Solutions Label Customization
    Ultralytics Solutions modules now let you show/hide labels and confidence scores, plus provide a unified label formatting method for a polished, consistent display.
    See PR by @RizwanMunawar

Improvements

  • YOLOE → YOLO11 Naming Consistency
    All references and documentation for YOLOE have been updated to align with YOLO11-based naming (like yoloe-11s-seg.pt), simplifying model selection and reducing confusion.
    See PR by @RizwanMunawar

  • Export Output Format Control
    Added explicit options to control bounding box output formats (xyxy vs. xywh) at export time, making integration clearer and more reliable.

  • Docker Image Upgrade
    The base Docker image now uses PyTorch 2.7.0 with CUDA 12.6 and cuDNN9 for improved runtime compatibility and performance.
    See PR by @glenn-jocher

  • Documentation Clarity & Developer Experience
    Improved return type annotations, refined docstrings, and updated code examples for a better developer onboarding and usage experience.
    See PR by @RizwanMunawar

Bug Fixes

  • More Reliable Non-Max Suppression (NMS)
    Fixed the class filtering logic in NMS for greater accuracy when filtering by object class, resulting in more dependable detection outputs.
    See PR by @RizwanMunawar
    See PR by @Y-T-G

  • Security Improvements
    Updated GitHub workflow permissions to read-only at the workflow level, in line with best CI/CD security practices.
    See PR by @glenn-jocher

:bullseye: Purpose & Impact

  • Greater Training Flexibility: Tackle class imbalance and hard examples with more configurable loss functions.
  • Consistent Model Selection: New YOLOE naming based on YOLO11 reduces confusion and boosts code reliability.
  • Performance Boosts: Half-precision exports and Docker updates lead to faster, more efficient deployments.
  • Robust Results: Upgraded NMS makes detections more accurate, especially with class filtering.
  • Clearer Visual Outputs & Docs: Enhanced label formatting, control options, and better documentation accelerate development.
  • Better Security: CI/CD improvements help safeguard the whole community.

:magnifying_glass_tilted_left: What’s Changed (Links & Authors)

:waving_hand: New Contributors

A warm welcome to:

Your contributions are much appreciated by the entire community!

:loudspeaker: Try It Out — We Value Your Feedback!

Update to Ultralytics v8.3.116 today to enjoy these enhancements.
Check out the full changelog for all the details.

We encourage you to try the new features, experiment with the improved models and training tools, and share your feedback in the community. Your insights are what drive Ultralytics forward!

Thank you for being an essential part of the YOLO ecosystem and helping us shape the future of vision AI.

Ultralytics Team (and the amazing YOLO community)