New Release: Ultralytics v8.3.214

Discussion
, ,
1

Ultralytics v8.3.214 :rocket: — Clearer classification visuals, sturdier training recovery, cleaner logs

Quick summary: v8.3.214 adds confidence scores to classification validation plots, hardens NaN recovery for more reliable training, resets results.csv on fresh runs, and updates docs to highlight SAM 3 and YOLO26. No breaking changes.

Highlights :glowing_star:

  • Classification validation plots now display top‑1 confidence for quicker debugging and clearer reads.
  • NaN/Inf recovery is more robust with safer train-mode handling and EMA recreation to prevent tensor errors.
  • New runs start with a fresh results.csv for tidy experiment management.
  • Documentation now surfaces SAM 3 (Coming Soon) and YOLO26 (Coming Soon), while keeping YOLO11 labels consistent.

Explore the release details in the v8.3.214 notes, which are available from the dedicated page for the Ultralytics v8.3.214 release, and check the full list of changes in the compare view from v8.3.213 to v8.3.214.

New Features :sparkles:

  • Classification visuals
    • Show top‑1 confidence on classification prediction plots during validation for clearer insights, as implemented in PR #22365 by @Y-T-G.

Improvements :hammer_and_wrench:

  • Training stability and recovery

    • Ensure the model is in train mode before NaN recovery checkpoint load to avoid inference‑tensor issues, introduced in PR #22381 by @glenn-jocher.
    • Recreate EMA before loading EMA state to prevent InferenceTensor errors when resuming, added in PR #22372 by @Y-T-G.
    • Improve NaN/fitness collapse detection by checking self.loss instead of self.tloss, refined in PR #22382 by @glenn-jocher.

  • Experiment logging

    • Reset results.csv when not resuming so new runs don’t append to old logs with exist_ok=True, adjusted in PR #22364 by @Y-T-G.

Documentation and Navigation :books:

  • SAM 3 docs page (Coming Soon), YOLO26 marked “Coming Soon,” and YOLO11 label cleanup for consistency; site navigation updated in PR #22373 by @glenn-jocher.
  • Learn about our recommended default model in the overview for YOLO11 models, and preview what’s next with the YOLO26 roadmap page.

Try the update :high_voltage:

Upgrade in seconds to get the latest improvements:

pip install -U ultralytics

See the new classification confidence overlays during validation with YOLO11:

yolo classify val model=yolo11n-cls.pt data=path/to/dataset

Or in Python:

from ultralytics import YOLO

model = YOLO('yolo11n-cls.pt')
metrics = model.val(data='path/to/dataset')  # validation plots now show top-1 confidence

Note: YOLO11 is our latest stable and recommended model for all use cases.

What’s Changed :clipboard:

Browse all changes in the v8.3.214 changelog comparison.

Feedback welcome :speech_balloon:

We’d love your thoughts on this release. Share feedback or questions in Ultralytics Discussions, or report issues directly by choosing to open a GitHub issue. Your input helps us keep YOLO fast, reliable, and easy to use—huge thanks to the community and contributors who made v8.3.214 possible!