New Release: Ultralytics v8.4.28

Discussion
, ,
1

Ultralytics v8.4.28 is here :rocket:

Quick summary: Ultralytics v8.4.28 makes training on small datasets more reliable with smarter autobatch behavior, while also improving network robustness, backend efficiency, and docs/test stability. If you train, validate, or deploy with Ultralytics YOLO, this release should feel smoother in a lot of everyday workflows. :sparkles:

:glowing_star: Highlights

:package: Smarter autobatch for small datasets

The biggest improvement in this release is that autobatch now respects dataset size, helping avoid unrealistic batch recommendations when training on tiny datasets.

From PR #24020 by @glenn-jocher:

  • Added dataset_size handling through Trainer.auto_batch(), check_train_batch_size(), and autobatch()
  • Skips profiling candidate batch sizes larger than the dataset
  • Clamps the final recommended batch size as an extra safety check

Why it matters: better defaults, less setup friction, and fewer odd edge cases when training on smaller custom datasets. :white_check_mark:

:globe_with_meridians: More reliable networking for datasets and Platform

We also improved retry behavior for transient networking issues across dataset downloads and Platform URL resolution.

From PR #24010 by @glenn-jocher:

  • Smarter retries for timeouts, connection drops, rate limits, and server-side failures
  • Faster failure on non-retryable issues like local save errors
  • Increased Platform URL resolution retries from 3 to 5
  • Clearer backoff and retry logging

Why it matters: more resilient cloud and data workflows, especially in CI or unstable network environments. :shield:

Improvements

:high_voltage: TensorFlow backend import optimization

From PR #24017 by @Y-T-G:

  • TensorFlow is now imported only for saved_model and pb formats in the TensorFlow backend loader
  • Avoids unnecessary dependency loading for unrelated inference paths

Impact: cleaner runtime behavior, reduced startup overhead, and fewer avoidable environment issues. :broom:

:blue_book: Ultralytics Platform docs refresh

From PR #24014 by @mykolaxboiko:

  • Removed outdated Activity tab references from model docs
  • Expanded onboarding documentation to cover tours and restart options

Impact: a smoother experience for users getting started with Ultralytics Platform, especially as it continues to be the simplest way to annotate datasets, train, and deploy YOLO models.

Bug Fixes

:test_tube: More reliable YOLO26 OBB CI testing

From PR #24011 by @glenn-jocher:

  • Updated the boats.jpg test image URL to a fixed jsDelivr CDN path for more consistent test behavior

Impact: improved CI stability and fewer flaky test failures. :wrench:

Why upgrade?

With v8.4.28, you can expect:

  • Better small-dataset training defaults
  • More robust network-dependent workflows
  • Leaner backend imports
  • Improved test and documentation reliability

If you’re starting a new project, we recommend using YOLO26, our latest stable and recommended model family. YOLO11 remains fully supported, but YOLO26 is the best choice for new work. :bullseye:

Try it out

Upgrade with:

pip install -U ultralytics

You can review the full release on the v8.4.28 release page, and browse everything included in the full changelog.

Feedback welcome :speech_balloon:

Please give v8.4.28 a try and let us know how it works for your training, inference, and deployment workflows. Feedback, edge cases, and regression reports are always appreciated—they help the whole Ultralytics community move faster.