New Release: Ultralytics v8.4.36

Discussion
, ,
1

Ultralytics v8.4.36 is out! :rocket:

This is a stability-focused release that restores reliable checkpoint-based training workflows and tightens up several docs across Ultralytics YOLO and Ultralytics Platform. The biggest win in v8.4.36 is a targeted training fix that makes checkpoint reuse dependable again, especially for Platform-based workflows. :white_check_mark:

Highlights

  • :brain: Critical training regression fixed for checkpoint-backed workflows
  • :test_tube: Regression test coverage added to protect the fix going forward
  • :books: Docs improved for Platform, Explorer, and Jetson/DeepStream guidance
  • :bullseye: Better reliability and clarity with minimal behavior changes

New Features & Major Improvements

:brain: Reliable checkpoint-based training is back

The most important change in v8.4.36 is the fix in Model.train() from PR #24167 by @glenn-jocher.

This update restores checkpoint-backed model seeding so that when a .pt model is already loaded, training reuses that in-memory model directly instead of incorrectly depending on trainer.args.model, which may not be a file path in Platform workflows.

Why this matters:

  • :white_check_mark: Makes checkpoint-based training reliable again
  • :white_check_mark: Prevents confusing failures in Platform-style flows
  • :white_check_mark: Keeps the fix narrowly scoped to the regression introduced earlier
  • :white_check_mark: Adds a dedicated regression test so this path stays protected

Improvements

:test_tube: Added regression coverage

Alongside the training fix in PR #24167 by @glenn-jocher, this release adds test coverage to confirm that training reuses the loaded checkpoint model instead of re-parsing the model source.

Additional validation also confirmed healthy behavior across:

  • detect
  • segment
  • classify
  • core Python model methods

:books: Platform documentation alignment

Documentation updates in PR #24157 by @glenn-jocher improve consistency around Ultralytics Platform, including:

  • required integration version now consistently documented as ultralytics>=8.4.35
  • clearer deployment defaults such as scale-to-zero behavior
  • single active instance cap clarification
  • account route auth scope updates
  • clearer dataset/version/export flow
  • improved training and checkpoint expectations

If you’re using Platform to annotate, train, and deploy models, these updates should make setup and behavior much easier to understand.

Bug Fixes

:magnifying_glass_tilted_right: Explorer docs version requirement corrected

PR #24165 by @raimbekovm updates the Explorer docs to clarify that Explorer was removed in >=8.3.12, with guidance to pin 8.3.11 if needed.

:desktop_computer: Jetson DeepStream archived links fixed

PR #24162 by @Y-T-G updates Jetson DeepStream references so older JetPack versions now point to the correct NVIDIA archived documentation.

Why upgrade?

v8.4.36 is especially valuable if you:

  • train from loaded .pt checkpoints
  • use Ultralytics Platform workflows
  • want lower regression risk in training
  • rely on up-to-date deployment and integration docs

In short: more reliable training, clearer docs, fewer surprises. :raising_hands:

Try it today

Upgrade with:

pip install -U ultralytics

Then verify your version with:

yolo version

You can also explore the release in the full release page for v8.4.36, and review everything included in the full changelog from v8.4.35 to v8.4.36.

Thank you :blue_heart:

Big thanks to the contributors who helped ship this release:

Please give v8.4.36 a try and let us know how it works for you — feedback, testing, and bug reports from the community help keep Ultralytics YOLO improving with every release.