New Release: Ultralytics v8.4.112

Discussion
, ,
1

:rocket: Ultralytics v8.4.112: Clearer, More Reliable Model Exports

Ultralytics v8.4.112 improves export reliability and makes deployment compatibility easier to understand. This release adds comprehensive task-support documentation across export formats, fixes DEEPX classification exports, expands validation coverage, and removes outdated TensorFlow GraphDef benchmark restrictions. :package::white_check_mark:

:books: New Export Support Documentation

Supported-task tables have been added to 20 integration pages, clearly documenting compatibility for all seven Ultralytics tasks:

  • Object detection
  • Instance segmentation
  • Semantic segmentation
  • Pose estimation
  • OBB detection
  • Classification
  • Depth estimation

The new tables also identify model-family constraints and unsupported combinations. For example, semantic segmentation and depth estimation are YOLO26-only for many formats, while specific limitations are now clearly documented for Axelera, Hailo, and Sony IMX500.

This should make selecting the right export format faster and help prevent avoidable deployment failures. :compass:

:magnifying_glass_tilted_left: Broader Export Validation

Export and inference behavior was verified through 77 local checks across seven tasks, covering:

  • TorchScript
  • ONNX
  • OpenVINO
  • CoreML
  • TensorFlow formats
  • PaddlePaddle
  • MNN
  • NCNN
  • ExecuTorch
  • LiteRT

TensorRT and RKNN support was additionally confirmed through existing CI test matrices, providing greater confidence that the documented compatibility matches real-world behavior.

:hammer_and_wrench: Bug Fixes

DEEPX Classification Export

DEEPX classification exports now correctly discover calibration images. Classification datasets store their image directory under root rather than img_path, and the exporter now handles this structure correctly.

DEEPX smoke tests were also expanded to cover every task-specific default model instead of testing only YOLO26 detection.

TensorFlow GraphDef Benchmarks

Obsolete restrictions that rejected OBB and pose models during TensorFlow GraphDef benchmarks have been removed. Supported OBB and pose workflows can now be benchmarked correctly. :high_voltage:

:sparkles: Additional Improvements

  • Updated the package version to 8.4.112.
  • Improved documentation wording for Huawei Ascend, TensorFlow GraphDef, NCNN, RKNN, and Hailo.
  • Clarified that supported Edge TPU workflows may still execute unsupported operations on the CPU.
  • Made export compatibility and model-family limitations easier to identify at a glance.

:raising_hands: Community Contribution

The release is driven by PR #25511 from @glenn-jocher, which documents supported tasks across every export format.

:package: Get Started

Upgrade to the latest release with:

pip install -U ultralytics

You can review the Ultralytics v8.4.112 release for the complete release details or explore the full comparison from v8.4.111 to v8.4.112 to inspect every change.

Give the new release a try, especially if you use model exports, DEEPX classification, or TensorFlow GraphDef benchmarking. We’d love to hear your feedback and learn about any export workflows that could be improved further! :rocket: