New Release: Ultralytics v8.3.128

:rocket: Ultralytics v8.3.128 Release: Enhanced Tracking, Multi-GPU Reliability, and Global Docs!

:glowing_star: Summary

We’re thrilled to introduce Ultralytics v8.3.128! This release brings major strides in object tracking reliability, multi-GPU training robustness, platform compatibility, and a significantly improved global user experience through extended documentation and easier navigation. Dive in to see how these updates streamline your workflows and empower users worldwide!


:sparkles: New Features & Improvements

Object Tracking & Re-Identification (ReID) Enhancements

  • Expanded test coverage of object tracking, including more GMC (Global Motion Compensation) methods and broader ReID model support.

    PR #20499 by @Y-T-G

  • Simplified and strengthened feature extraction for trackers, especially benefiting BoT-SORT compatibility.
  • Improved tracker initialization and robust feature handling for better, more reliable results.

Multi-GPU Training & Device Handling

  • Added strong checks and assertions for distributed (multi-GPU/Distributed Data Parallel) training to prevent CUDA and device selection mishaps.

    PR #20516 by @Y-T-G
    PR #20502 by @Y-T-G

  • Improved device selection logic for reliable post-training validation.

Platform Compatibility

  • PaddlePaddle model export is now clearly blocked on NVIDIA Jetson devices with informative errors, reducing confusion.

    PR #20490 by @lakshanthad

  • Enhanced support for TensorRT inference on Jetson by accommodating Python versions up to 3.8.10 and patching NumPy issues.

    PR #20485 by @lakshanthad

VisualAISearch & CLIP Integration

  • VisualAISearch now leverages Ultralytics’ internal CLIP implementation, reducing external dependencies and improving compatibility.

    PR #20513 by @RizwanMunawar

  • Documentation for VisualAISearch updated with comprehensive parameter details for easier onboarding.

    PR #20507 by @RizwanMunawar

Documentation & Internationalization


:hammer_and_wrench: Bug Fixes & Polishing


:bullseye: Purpose & Impact

  • Greater Reliability: More robust tracking and multi-GPU handling benefit advanced users and large deployments.
  • Better Compatibility: Smoother experience on Jetson, Raspberry Pi, and across Python versions expands deployment versatility.
  • Simplified Setup: Reduced third-party dependency for VisualAISearch and clearer docs help users onboard and troubleshoot.
  • Global Accessibility: Multilingual docs and easier navigation reach and empower users globally.
  • User Guidance: Improved errors and documentation help prevent and solve common issues quickly.

:inbox_tray: Try It & Share Your Thoughts!

Give Ultralytics v8.3.128 a spin—whether you’re training on multiple GPUs, deploying on edge devices, or exploring our enriched documentation experience, we’d love to hear from you!
Your feedback guides our future improvements—please share your experiences, ideas, and questions right here in this thread.

Full changelog and details:
Ultralytics v8.3.128 comparison


A big thank you to all contributors, users, and the dedicated Ultralytics team that make this possible. As always, the YOLO community’s feedback and participation drive our progress forward. Happy building and exploring!