New Release: Ultralytics v8.4.58

Discussion
, ,
1

Ultralytics v8.4.58 is out :rocket:

We’ve released Ultralytics v8.4.58, a reliability-focused update that makes model export testing more stable, isolated, and maintainable across a wide range of deployment targets. While this release does not introduce a new model architecture, it delivers important behind-the-scenes improvements that strengthen export support for Ultralytics YOLO users working across mobile, embedded, browser, and accelerator-specific environments.

If you rely on exporting models for deployment, this release should give you more confidence that formats are being tested in ways that better reflect real-world usage :white_check_mark:

Quick Summary

  • :hammer_and_wrench: Redesigned export testing around isolated environments
  • :puzzle_piece: Added environment metadata to each export format
  • :test_tube: Introduced --export-env for cleaner test selection
  • :gear: Added a new IsolatedExports CI workflow
  • :package: Split export dependencies into smaller targeted extras
  • :white_check_mark: Added validation checks for export registry consistency
  • :robot: Added RKNN export test coverage
  • :spouting_whale: Improved Docker cleanup for export images

You can explore the full release on the Ultralytics v8.4.58 release page and review all changes in the full changelog from v8.4.57 to v8.4.58.

New Features

:vertical_traffic_light: Isolated export testing in CI

The biggest change in this release is that export tests are now organized into dedicated environments instead of being run together in one shared CI setup.

This work came in PR #24649 by @glenn-jocher.

Why this matters:

  • Reduces dependency conflicts between export backends
  • Better reflects how users actually install and use specific export targets
  • Makes specialized formats easier to test and maintain over time

This is especially helpful for formats with conflicting or heavyweight dependencies, including TensorFlow, CoreML, MNN, NCNN, ExecuTorch, IMX, RKNN, Axelera, and DEEPX.

:puzzle_piece: Export formats now declare their own environment

The internal export_formats() registry now includes environment metadata for every export target. This allows CI to understand exactly where each export should run, rather than relying on fragile manual logic.

:test_tube: New --export-env test selection

Tests can now be selected by export environment using --export-env, replacing older skip-based patterns with something much cleaner and less error-prone.

Improvements

:gear: New IsolatedExports workflow

A new IsolatedExports CI job now creates dedicated virtual environments and runs smoke tests plus export tests inside them. This replaces the older shell-script-based coordination and makes the pipeline easier to reason about and extend.

:package: Leaner dependency groups

Export dependencies were split into smaller extras, helping keep installs more targeted and avoiding unnecessary heavy packages in base workflows.

Examples include:

  • export-base
  • export-tensorflow
  • export-coreml
  • export-executorch
  • export-deepx
  • export-legacy-torch

This should help both CI efficiency and local development workflows.

:globe_showing_europe_africa: Better support for specialized deployment targets

This release improves validation paths for hardware-specific and edge deployment formats such as:

  • IMX
  • RKNN
  • Axelera
  • DEEPX
  • ExecuTorch

That means stronger coverage for users deploying Ultralytics YOLO models beyond standard desktop or server environments.

Bug Fixes and Reliability Enhancements

:white_check_mark: Added checks for export registry consistency

New tests now verify that:

  • every export environment has a smoke test
  • every export format points to a valid registered environment

These checks help prevent registry mismatches and reduce maintenance mistakes as export support expands.

:robot: RKNN export test added

A dedicated RKNN export test was added, improving automated coverage for Rockchip deployment workflows.

:spouting_whale: Docker cleanup improvements

Export Docker images now clean caches more aggressively, reducing leftover build artifacts and helping keep workflows tidier.

Why this release matters

Even though v8.4.58 is not a major end-user feature release, it has a meaningful impact for production and deployment workflows:

  • :locked: More reliable exports across formats
  • :rocket: Faster and leaner CI for routine development
  • :broom: Simpler maintenance for export support going forward
  • :pushpin: Lower risk of one export backend breaking another

In short: this release improves the foundation that powers model export across the ecosystem.

Try it out

To upgrade locally:

pip install -U ultralytics

If you’re starting a new project, we recommend building with Ultralytics YOLO26, our latest stable model family. You can also manage datasets, training, export, and deployment workflows through the Ultralytics Platform, which is the simplest way to annotate, train, and deploy YOLO models.

PRs in this release

Feedback

Please give v8.4.58 a try and let us know how it works for your export workflows :raising_hands:
If you spot any regressions, edge cases, or deployment-specific issues, we’d love your feedback from the community.