New Release: Ultralytics v8.4.54

Discussion
, ,
1

Ultralytics v8.4.54 is here :rocket:

Quick summary: Ultralytics v8.4.54 brings a major deployment upgrade with local Qualcomm QNN export for Snapdragon devices :mobile_phone::high_voltage:, plus cleaner export docs, imgsz support across all Solutions, improved tuning compatibility, and several reliability fixes for segmentation and image caching.

If you’re building with Ultralytics YOLO, especially for edge and mobile deployment, this is a great release to try.

:glowing_star: Highlights

:mobile_phone: New: Qualcomm QNN export for Snapdragon devices

The headline feature in v8.4.54 is a new qnn export format, making it much easier to deploy YOLO models on Qualcomm Snapdragon hardware.

With format="qnn", you can now export a model locally into a QNN-ready package using ONNX Runtime’s QNN Execution Provider:

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
model.export(format="qnn")

Why this matters:

  • :white_check_mark: No Qualcomm account required
  • :white_check_mark: No cloud upload required
  • :white_check_mark: No separate Qualcomm SDK download required
  • :white_check_mark: Supports Snapdragon deployment paths including CPU, GPU, and Hexagon NPU

This is a big step forward for mobile, embedded, automotive, and low-power AI workflows.

:electric_plug: Better QNN workflow integration

QNN is now integrated more cleanly into the standard Ultralytics export and backend flow:

  • Registered as a supported export target
  • Dedicated backend support for inference and validation workflows
  • Automatic disabling of end-to-end export where QNN does not support it, helping avoid invalid configurations

:blue_book: Export docs cleanup and standardization

Deployment documentation has been reorganized into a clearer Export / Predict / Validate structure across formats.

This release also adds a dedicated guide for Qualcomm QNN and improves the overall deployment experience across formats such as ONNX, TensorRT, OpenVINO, NCNN, TFLite, and CoreML. You can explore the latest docs in the Ultralytics documentation.

Improvements

:framed_picture: imgsz support added to all Solutions

All Solutions now support the imgsz argument, giving users more direct control over inference input size and the speed/accuracy tradeoff. Logging was also improved so reported input shapes better match actual model input.

:bullseye: Ray Tune compatibility update

Tuner trial ID handling was updated for compatibility with Ray 2.7+, while keeping support for older versions.

:eyes: Docs updates for validation and prediction

A couple of nice usability improvements landed in the docs:

  • Validation docs now include show_labels and show_conf
  • Prediction docs now explicitly list .jpeg2000 as a supported image format

Bug Fixes

:hammer_and_wrench: Segmentation and data-loading reliability

This release includes a few important fixes to make training and inference more robust:

  • Fixed semantic segmentation mask loading when masks appear as (H, W, 1)
  • Fixed stale .npy cache issues that could cause grayscale models to load RGB data unexpectedly
  • Improved dtype consistency in semantic segmentation loss, especially helpful for mixed-precision training

:brain: Tracking utility optimization

Some SciPy usage in tracking internals was replaced with NumPy equivalents, helping simplify dependencies and improve maintainability.

What changed

Why upgrade?

Ultralytics v8.4.54 is especially valuable if you:

  • Deploy on Qualcomm Snapdragon
  • Want a smoother path from training to edge inference
  • Use Solutions and need configurable imgsz
  • Run segmentation pipelines with mixed or unusual mask formats
  • Depend on Ray Tune in newer environments

For new projects, we recommend starting with Ultralytics YOLO26, our latest stable model family. If you’re managing datasets, training, and deployment workflows, the Ultralytics Platform remains the simplest way to get started.

Try it out :raising_hands:

You can upgrade with:

pip install -U ultralytics

See the v8.4.54 release page for the full release, and browse the full changelog for all details.

We’d love your feedback — especially if you test the new QNN export flow on Snapdragon devices.