New Release: Ultralytics v8.4.62

Discussion
, ,
1

:rocket: Ultralytics v8.4.62 Released

:glowing_star: Summary

We’ve released Ultralytics v8.4.62, a high-value reliability update focused on one of the most important parts of training: making sure your trained models are actually saved :floppy_disk:

The headline fix prevents successful training runs from ending without checkpoints due to EMA NaN/Inf issues during save checks, especially in some AdamW + AMP workflows. This release also includes a substantial Ultralytics Platform docs refresh, improved test stability, cleaner external documentation links, and leaner CI.

If you train Ultralytics YOLO models regularly, this is a strongly recommended upgrade :white_check_mark:

:fire: Key Highlights

:floppy_disk: Major training fix: checkpoints are no longer discarded

The most important change in this release comes from PR #24731 by @glenn-jocher, with author profile @glenn-jocher.

This fixes a bug where valid training runs could complete successfully but still fail to save any checkpoint.

This was especially relevant for some setups using AdamW + AMP, where validation could corrupt live EMA weights and trigger repeated warnings like Skipping checkpoint save... EMA contains NaN/Inf.

:white_check_mark: Safer validation during AMP training

With the fix in PR #24731 by @glenn-jocher, validation still benefits from mixed precision speedups, but it no longer permanently modifies or “poisons” the live EMA model.

That means:

  • validation no longer mutates the live EMA in place
  • finiteness checks now happen on the original fp32 EMA
  • overflow is safely clamped during checkpoint serialization instead of skipping the save

:test_tube: New coverage for overflow-safe checkpoint saving

This release also adds test coverage to ensure models with large-but-finite EMA weights are still saved correctly, helping protect against regressions in future updates.

:blue_book: Documentation Improvements

:books: Big Ultralytics Platform docs refresh

The documentation update in PR #24726 by @glenn-jocher, from @glenn-jocher, significantly improves accuracy and API coverage across the Ultralytics Platform documentation.

Highlights include:

  • corrected UI labels and workflows
  • expanded Platform API reference
  • clearer docs for datasets, annotations, training, deployment, billing, teams, and integrations
  • newly documented API capabilities such as dataset embeddings, class management, GPU availability, and import flows

If you use the Ultralytics Platform, this should make day-to-day workflows much easier to follow :raising_hands:

:link: Fixed broken COCO evaluation links

In PR #24723 by @glenn-jocher, from @glenn-jocher, outdated COCO evaluation links were replaced with the canonical COCO upload instructions, and old docs author profile links were corrected.

:gear: Improvements and Maintenance

:test_tube: Less flaky data-related tests

PR #24724 by @glenn-jocher, authored by @glenn-jocher, reduces unnecessary downloads in tests and reuses cached assets when possible.

This should improve CI dependability and reduce false failures.

:high_voltage: Leaner CI workflows

Two updates helped streamline CI:

:bullseye: Why this release matters

Ultralytics v8.4.62 is not a major model-feature release, but it is a very important stability release :hammer_and_wrench:

It helps by:

  • preventing trained YOLO models from being lost at the end of training
  • improving confidence in AMP-based training workflows
  • reducing wasted compute from failed checkpoint saves
  • making Platform docs more accurate and easier to use
  • improving internal test and CI reliability

If you’ve ever seen a run finish well but end with no checkpoint saved, this release directly addresses that issue.

:package: What’s Changed

You can also review the full changelog for v8.4.62 or browse the release page for v8.4.62.

:raising_hands: Try it out

We recommend upgrading and giving it a spin, especially if you train with AMP or AdamW:

pip install -U ultralytics

If you’re starting a new project, we recommend Ultralytics YOLO26 for the best current performance and end-to-end support across detect, segment, classify, pose, and OBB.

:speech_balloon: Feedback

Please try v8.4.62 and let us know how it works for your training and deployment workflows. Feedback, bug reports, and regression checks are always appreciated — they help us keep YOLO reliable for everyone.