New Release: Ultralytics v8.4.71

Discussion
, ,
1

Ultralytics v8.4.71 is out :tada:

Quick summary: Ultralytics v8.4.71 is a small release on the surface, but it packages a major upgrade for C++ deployment workflows. The biggest highlight is a full refactor and unification of the official C++ inference examples, making it much easier to deploy Ultralytics YOLO models across multiple backends and tasks. We also added new Blackwell GPU options to the Ultralytics Platform docs :high_voltage:

Highlights

:hammer_and_wrench: Unified C++ examples for easier deployment

This release brings a substantial cleanup and reorganization of the official C++ examples into a single examples/cpp/ structure.

That means a more consistent experience across backends including:

  • OpenCV-DNN
  • ONNX Runtime
  • LibTorch
  • MNN
  • OpenVINO
  • Triton

Shared header-only utilities now cover:

  • detection
  • post-processing
  • rendering and annotation
  • CLI handling
  • COCO fallback names

For C++ users, this is the real headline of v8.4.71 :rocket:

:bullseye: Broader task support across C++

The refactored examples now support all major YOLO tasks in C++:

  • detection
  • segmentation
  • pose
  • OBB
  • classification
  • YOLO26 semantic segmentation

They also support multiple model generations, including YOLOv8, YOLO11, and YOLO26, our latest recommended model for new projects.

:robot: Automatic task detection

Many of the updated C++ examples can now automatically infer the task from model metadata or output shapes. This reduces manual setup and makes it easier to swap exported models without reworking your pipeline.

:toolbox: More consistent developer experience

The new C++ layout introduces a cleaner and more standardized workflow, including:

  • unified build targets like yolo_<backend>
  • standardized CLI args such as --model and --source
  • cleaner READMEs and improved docs
  • lightweight YOLO26 defaults in common examples

:cloud: Expanded Platform GPU docs

We also expanded the Ultralytics Platform documentation with two new Blackwell GPU options:

  • RTX PRO 4000 Blackwell with 24 GB VRAM
  • RTX PRO 5000 Blackwell with 48 GB VRAM

These updates help Platform users better choose the right cloud hardware for training and deployment workloads.

Version bump

This release includes the package update from 8.4.70 to 8.4.71 :package:

While the version-bump PR itself does not introduce direct Python API or model behavior changes, it packages the C++ improvements and documentation updates into the latest official release.

What changed

New Features and Improvements

Why this matters

For most users

v8.4.71 is primarily a packaging and release update that keeps environments aligned with the latest Ultralytics build :white_check_mark:

For C++ developers

This release significantly lowers the friction of deploying YOLO in production C++ systems. Instead of learning different example layouts for each backend, you now get a more unified workflow with broader task support.

For teams adopting YOLO26

This release strengthens deployment support for YOLO26, including semantic segmentation, making it easier to bring the latest recommended model into real-world applications.

Try it out

If you want to upgrade locally:

pip install -U ultralytics

You can explore the full release in the v8.4.71 release page, and review everything included in the full changelog from v8.4.70 to v8.4.71.

Thanks as always to the community and contributors who help keep Ultralytics moving forward :blue_heart:

We’d love for you to try v8.4.71, especially the new C++ example structure, and share your feedback in the discussion!