New Release: Ultralytics v8.4.50

Discussion
, ,
1

Ultralytics v8.4.50 is here :rocket:

Summary: Ultralytics v8.4.50 is a deployment-focused release centered on new DeepX export and inference support for running Ultralytics YOLO models on DeepX NPU edge hardware. This update also improves tuning reliability, fixes mixed-precision fusion issues, and adds clearer RT-DETR latency tuning guidance.

:glowing_star: Highlights

DeepX export support for edge deployment :brain::high_voltage:

The biggest addition in v8.4.50 is DeepX export integration, which lets you export models directly with format="deepx" as part of the standard Ultralytics workflow.

With DeepX export, Ultralytics now:

  • converts the model through ONNX and compiles it to DeepX .dxnn
  • creates a complete deployment bundle including:
    • compiled model file
    • config.json for preprocessing and calibration
    • metadata.yaml for model metadata
  • enables int8=True automatically for INT8-focused deployment
  • defaults to coco128.yaml for calibration if no dataset is provided
  • supports modern tasks including detect, segment, classify, pose, and OBB

This makes it much easier to move YOLO models onto low-power embedded AI systems such as smart cameras, robotics platforms, and industrial devices.

DeepX inference backend added :electric_plug:

DeepX models exported from Ultralytics can now be loaded back for inference through a dedicated backend, making the deployment workflow more integrated end to end.

DeepX docs and references added :blue_book:

This release also includes documentation updates to help users get started with the new workflow more quickly, including guides, reference pages, and export documentation improvements.

Improvements :hammer_and_wrench:

More honest hyperparameter tuning results :white_check_mark:

Hyperparameter tuning now reports failures more accurately:

  • failed runs are reflected in status output
  • misleading success reporting is removed when all iterations fail
  • best_hyperparameters.yaml is no longer written if no valid result exists

This should make tuning output much more trustworthy for experimentation and production workflows.

Better mixed-precision fusion stability :bullseye:

A dtype mismatch bug in fuse_conv_and_bn and fuse_deconv_and_bn has been fixed, improving reliability when working with float16 or bfloat16.

Docs updates :books:

RT-DETR eval_idx guidance for lower latency :stopwatch:

The RT-DETR docs now explain how to use eval_idx to reduce decoder layers at inference time, trading a small amount of accuracy for lower latency without retraining.

Important notes :warning:

  • DeepX export currently requires x86-64 Linux
  • ARM64 export is not supported
  • inference support depends on the DeepX runtime and dx_engine

What changed in GitHub

You can also explore the full changelog for v8.4.50 or view the official v8.4.50 release.

Try it out :raising_hands:

If you’re targeting edge hardware, this is a great release to test. For example:

from ultralytics import YOLO

model = YOLO("yolo11n.pt")
model.export(format="deepx")

If you’re starting a new project, we recommend using Ultralytics YOLO26 on Ultralytics Platform for the latest models and workflows, while YOLO11 remains fully supported.

Please give v8.4.50 a try and let us know how it works for your deployment, training, or optimization workflows. Feedback, issues, and benchmarks are always welcome :blue_heart: