New Release: Ultralytics v8.4.41

Discussion
, ,
1

:tada: Ultralytics v8.4.41 is out!

Quick summary: Ultralytics v8.4.41 delivers a major SAM3 video tracking quality fix to reduce ghost IDs :ghost:, a safer NDJSON dataset conversion pipeline for parallel workflows :card_index_dividers:, and a broad set of documentation and Ultralytics Platform improvements to make training, deployment, and customization easier :rocket:

If you’re building with Ultralytics YOLO, this is a strong quality-of-life release with especially useful updates for video tracking and data pipeline reliability.

:glowing_star: Highlights

:bullseye: Better SAM3 video tracking quality

The biggest update in v8.4.41 is the tracking fix in PR #24249 by @Y-T-G, with contributor profile @Y-T-G.

This change improves SAM3 video tracking by:

  • Enabling masklet confirmation filtering by default
  • Reducing false positive “ghost” object IDs
  • Tightening tracker keep-alive settings:
    • init_trk_keep_alive: 30 → 10
    • max_trk_keep_alive: 30 → 10
  • Improving handling of unconfirmed IDs in single-frame inference output

:card_index_dividers: Safer NDJSON dataset conversion

The data pipeline also got a nice reliability upgrade in PR #24290 by @glenn-jocher, with author profile @glenn-jocher.

Key improvements include:

  • Hash-qualified output folders for converted datasets
  • Better protection against cache collisions during concurrent jobs
  • Hashing that ignores temporary signed-URL query strings
  • More robust class-count inference when class_names are missing

This should make repeated conversions and team-scale workflows much more reliable.

:rocket: New Guides and Documentation Improvements

This release includes a strong docs refresh for both YOLO users and teams using the Ultralytics Platform, available in the Platform docs and on the Ultralytics Platform.

:books: New and improved guides

These updates should make it easier to:

  • Fine-tune YOLO26 on your own data
  • Reuse configuration files with cfg
  • Explore faster preprocessing pipelines
  • Get started with serverless deployment
  • Navigate Platform features with less friction

:test_tube: CI and Benchmark Stability

For contributors and users watching CI closely, benchmark stability was improved in PR #24286 by @lakshanthad, with author profile @lakshanthad.

Changes include:

  • Moving benchmark runners from ubuntu-latest to cpu-latest
  • Slight OBB threshold tuning to reduce flaky benchmark failures

:wrench: Why this release matters

v8.4.41 is a focused release, but the impact is meaningful:

  • :white_check_mark: Cleaner video tracking with fewer phantom IDs in SAM3 workflows
  • :high_voltage: Faster stale-track cleanup for more responsive tracking behavior
  • :locked: Safer dataset conversion in parallel and repeated NDJSON workflows
  • :blue_book: Better onboarding and reference docs for advanced workflows
  • :globe_with_meridians: Clearer Ultralytics Platform guidance for production usage

:package: Upgrade now

You can update with:

pip install -U ultralytics

Then verify your install with:

yolo version

You can explore the full release on the v8.4.41 release page, and browse every commit in the full changelog from v8.4.40 to v8.4.41.

:speech_balloon: Try it out and let us know what you think

If you’re using SAM3 tracking, NDJSON datasets, or the latest Ultralytics YOLO26 workflows, give v8.4.41 a spin and share your feedback. Community testing and reports help us keep improving the release experience for everyone :raising_hands:

Thanks to all contributors who helped make this release happen!