New Release: Ultralytics v8.4.100

Discussion
, ,
1

:rocket: Ultralytics v8.4.100: Hailo Segmentation, Stronger Exports, and Better Reliability

Ultralytics v8.4.100 expands edge deployment with Hailo instance segmentation support for YOLOv8 and YOLO11, clearer INT8 and TensorRT guidance, broader fuzz testing, improved Platform workers, cleaner training errors, and more reliable Axelera exports.

Explore the v8.4.100 release for the complete release details.

:glowing_star: New Features

:green_circle: Hailo Instance Segmentation

PR #25259 by @JESUSROYETH adds Hailo HEF export and inference support for YOLOv8-seg and YOLO11-seg.

  • Supports Hailo-8, Hailo-8L, Hailo-10, and Hailo-15 accelerators.
  • Exports raw segmentation tensors for Ultralytics-side decoding, bounding-box processing, NMS, and mask generation.
  • Preserves the standard Ultralytics prediction format and existing mask workflows.
  • Avoids Hailo on-chip NMS configuration for segmentation models.
  • Clearly rejects YOLO26 segmentation because Hailo-8 and Hailo-8L cannot allocate its attention-based architecture.

YOLO26 remains the latest stable and recommended Ultralytics model for new projects, while YOLO11 continues to be fully supported.

:books: Deployment Improvements

Better Hailo INT8 Guidance

PR #25260 by @JESUSROYETH adds practical INT8 accuracy guidance for YOLOv8, YOLO11, and YOLO26 on Hailo-8L.

The updated guidance explains:

  • Why calibration data should closely match the production domain.
  • Why YOLO26 confidence scores may be approximately 0.05 lower on Hailo devices.
  • Why attention layers can be more sensitive to INT8 quantization on Hailo-8 and Hailo-8L.

Clearer TensorRT Compatibility Warnings

PR #25263 by @lakshanthad improves Platform export documentation with clearer TensorRT compatibility requirements.

Downloaded TensorRT engines must match the target GPU family, TensorRT version, CUDA runtime, and build environment. When environments differ, users are now directed to export the engine locally on the deployment device.

:test_tube: Testing and Reliability

Expanded Fuzz Testing

PR #25262 by @glenn-jocher broadens fuzz coverage across:

  • Tracking and cached video.
  • Malformed media and Unicode paths.
  • Additional model families.
  • CoreML on macOS.
  • Shell-safe reproducer commands.

Duplicate commands are now skipped, and the fuzzing budget increases from 285 to 300 minutes per run for more effective exploration.

More Reliable Axelera Exports

PR #25267 by @glenn-jocher fixes compatibility issues with Axelera Devkit 1.7.0 by pinning omnimalloc==0.5.0.

Runtime checks now validate the same allocator version before export, reducing environment-related failures.

:hammer_and_wrench: Training and Platform Improvements

Cleaner Errors for Invalid Small Training Configurations

PR #25264 by @glenn-jocher replaces a confusing low-level PyTorch BatchNorm failure with an actionable error.

When batch=1 is combined with an image size that is too small, training now fails early and recommends increasing the batch size or image size.

Improved Managed-Worker Integration

PR #25266 by @glenn-jocher improves managed-worker integration by:

  • Reusing existing Platform training callbacks.
  • Supporting configurable API endpoints through PLATFORM_API_URL.
  • Respecting the current SETTINGS["runs_dir"] value when creating training output paths.

For a streamlined workflow covering dataset annotation, training, deployment, and monitoring, try the Ultralytics Platform.

:package: Get Started

Upgrade to the latest release:

pip install -U ultralytics

Review every change in the v8.4.99 to v8.4.100 comparison, then try the new Hailo segmentation workflow and deployment improvements.

Thanks to the YOLO community and all contributors who helped make this release possible. Please share your results, questions, and feedback below! :raising_hands: