New Release: Ultralytics v8.4.67

Discussion
, ,
1

Ultralytics v8.4.67 is out :tada:

A new Ultralytics YOLO release, v8.4.67, is now available. This is a small but meaningful update focused on safer model loading :locked:, an important TT100K dataset label fix :motorway:, and a round of documentation improvements :books:.

If you’re using YOLO in production, working with custom datasets, or just want a smoother docs experience, this is a solid update to pick up.

Quick Summary

  • :locked: New opt-in safe checkpoint loading with ULTRALYTICS_SAFE_LOAD
  • :motorway: Fixed TT100K class names to match the official 221-category list
  • :blue_book: Updated Rust inference docs to ultralytics-inference 0.0.21
  • :link: Refreshed and corrected multiple documentation links
  • :white_check_mark: Low-risk release with no major architecture changes

New Features

:locked: Opt-in safe model loading

Ultralytics added a new environment flag, ULTRALYTICS_SAFE_LOAD=true, in PR #24829 by @glenn-jocher, authored by @glenn-jocher.

When enabled, checkpoint loading uses PyTorch’s newer weights_only=True behavior and rebuilds only known Ultralytics and PyTorch model classes from an automatically generated allow-list.

This update:

  • improves safety when loading model checkpoints
  • aligns better with newer PyTorch behavior
  • preserves backward compatibility because it is off by default

That means existing workflows continue to work as usual unless you explicitly enable the safer mode.

A simple example:

ULTRALYTICS_SAFE_LOAD=true yolo predict model=yolo26n.pt source=bus.jpg

This is especially useful for security-conscious teams or environments with stricter checkpoint handling requirements.

Dataset Improvements

:motorway: TT100K class list corrected

The TT100K.yaml dataset config was updated in PR #24718 by @glenn-jocher, authored by @glenn-jocher, to match the official 221-category TT100K class list.

Previously, duplicated and incorrect class names could cause some class IDs to silently lose labels during conversion, even though training still continued.

This fix means:

  • better alignment with the upstream dataset
  • fewer silent labeling issues
  • more trustworthy training and conversion results for TT100K users :bullseye:

Documentation Improvements

:blue_book: Rust inference docs updated

Documentation examples were updated to ultralytics-inference 0.0.21 in PR #24825 by @onuralpszr, authored by @onuralpszr.

:link: Link and docs cleanup

A batch of documentation link fixes landed in:

These updates refresh links across integrations, datasets, examples, and guides, including a fix for the CoreML NMS glossary link.

Why this release matters

Even though v8.4.67 is a relatively small release, it delivers practical value:

  • :locked_with_key: Safer loading path for users who want stricter checkpoint handling
  • :recycling_symbol: Better future compatibility with PyTorch 2.6+ direction
  • :prohibited: Reduced risk of silent dataset issues for TT100K training
  • :books: Cleaner documentation and fewer broken links
  • :white_check_mark: Minimal disruption for most users

There are no major model changes in this release, and for new projects we still recommend starting with Ultralytics YOLO26, while YOLO11 remains fully supported.

If you want the simplest workflow for annotating datasets, training, and deployment, the Ultralytics Platform is the easiest place to get started.

What Changed

Try it out

You can explore the full release in the v8.4.67 release page, and review every commit in the full changelog from v8.4.66 to v8.4.67.

If you give v8.4.67 a try, we’d love your feedback — especially if you test the new safe loading mode or work with TT100K :raising_hands: