New Release: Ultralytics v8.4.19

Discussion
, ,
1

:rocket: Ultralytics v8.4.19 is out!

TL;DR: Ultralytics v8.4.19 makes Ultralytics Platform training sync much more reliable (especially model tracking via model_id), plus improvements for SAM outputs, lightweight model stability, and clearer YOLO26 end-to-end docs. :white_check_mark:

If you’ve ever had a run where artifacts/checkpoints didn’t feel perfectly “attached” to the right Platform session—this release is for you. :bullseye:

:glowing_star: New Features / Highlights

:red_circle: Platform training model_id fix (most important)

Training callbacks now consistently pass model_id during uploads and training events, improving continuity, cancellation handling, and final result linking.

  • PR: Platform training pass model ID fix by @glenn-jocher
  • What changed:
    • Refactored callback state into a single trainer.platform context (instead of scattered trainer fields), including:
      • model_id
      • last upload time
      • cancellation status
      • console/system loggers
    • Tightened upload + event flow for better session continuity :white_check_mark:
    • Improved Platform URL handling when servers adjust model slugs (e.g., name auto-increment)

:brain: Improvements

:bullseye: SAM prediction cleanup (fewer duplicate overlaps)

Added NMS in SAM3SemanticPredictor postprocessing and feature-inference paths, reducing overlapping duplicates and keeping boxes/masks aligned.

:brick: PSA attention edge-case fix (small/lightweight configs)

Prevents zero attention heads in PSABlock for very small channel configs, improving robustness for lightweight/custom scales.

:blue_book: Clearer end-to-end docs for YOLO26 / YOLOv10

Clarifies that in end-to-end mode (YOLO26, YOLOv10), max_det and agnostic_nms are supported directly, and that end2end=False is mainly for traditional IoU-based NMS behavior.

:hammer_and_wrench: Bug Fixes & Maintenance

:gear: CI updates (artifact actions bumps)

:movie_camera: Docs media update

Updated the African Wildlife dataset tutorial embed.

:test_tube: Try it now

Update:

pip install -U ultralytics

Quick sanity check (Ultralytics YOLO):

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
model.predict("https://ultralytics.com/images/bus.jpg")

If you’re using Ultralytics Platform, this release should noticeably improve training run tracking—especially around model uploads and session linking. :satellite_antenna:

:link: Release links

Grab details from the v8.4.19 release page and browse the full v8.4.18…v8.4.19 changelog.

:speech_balloon: Feedback welcome

Please try v8.4.19 and share:

  • any Platform training edge-cases you still hit :puzzle_piece:
  • SAM prediction quality differences you notice :bullseye:
  • lightweight/custom model configs that benefit from the PSA fix :brick:

Thanks to everyone contributing—this pace is powered by the Ultralytics team and the wider YOLO community. :raising_hands: