New Release: Ultralytics v8.3.229

Discussion
, ,
1

:rocket: Ultralytics v8.3.229 Release – 3× Faster COCO Segmentation, Safer Environments, Better Docs

Ultralytics v8.3.229 is now live! This release focuses on much faster COCO instance segmentation validation (~3×), safer dependency and environment controls, more reliable multi‑GPU behavior, and a big documentation refresh around real‑world deployments (Jetson, datasets, integrations).

If you’re training or validating YOLO11-seg on COCO‑style datasets or running YOLO in production environments, this one is for you. :high_voltage::brain:

:glowing_star: Summary

  • :high_voltage: Up to 300% faster COCO instance segmentation validation via a new in‑repo RLE pipeline.
  • :shield: Tighter dependency bounds and environment checks for more predictable installs and exports.
  • :abacus: More robust multi‑GPU training & validation, especially on small datasets.
  • :books: Updated docs and guides for Jetson AGX Thor (JetPack 7), datasets, and deployment stacks.
  • :speech_balloon: Improved docs build & chat widget experience for easier in‑docs assistance.

You can review the full release notes in the v8.3.229 GitHub release page.

:sparkles: New Features & Performance

:high_voltage: 300% Faster COCO Segmentation Validation

Thanks to a major optimization by @Laughing-q in PR #22651, COCO instance segmentation validation is now up to 3× faster:

  • New native, optimized RLE encoding pipeline for masks.
  • Cleaner, more efficient mask scaling using ops.scale_masks with ratio_pad support.
  • Removal of per‑mask thread pools for simpler and more predictable behavior.

This especially benefits frequent validation runs for YOLO11 segmentation models on COCO‑like datasets. :racing_car:

:locked_with_key: Environment & Dependency Improvements

We’ve put a lot of work into making installs and exports more stable and reproducible:

  • @onuralpszr added upper bounds on core dependencies like numpy, torch, opencv, pillow, polars, etc. in PR #22701 to avoid surprise breakages from upstream updates.
  • @Y-T-G enforced IMX‑specific export guards in PR #22673, explicitly asserting python>=3.8<3.12 and torch<2.9 for Sony IMX exports with clear error messages.
  • The official Docker image now uses PyTorch 2.9.1, CUDA 12.8, cuDNN 9 via PR #22708 by @glenn-jocher.
  • Docker builds are simpler and more robust after removing the explicit pip install uv in favor of the base image tools, via PR #22709 by @glenn-jocher.

These changes should significantly reduce “it worked yesterday, broken today” issues when you reinstall or update environments. :shield:

:abacus: Multi‑GPU & CI Reliability

Multi‑GPU workflows and CI coverage were strengthened to catch edge cases earlier:

  • @Y-T-G fixed a distributed sampler edge case where batch_size could be larger than the dataset size, now safely falling back to batch_size=1 in PR #22714.
  • GPU CI is re‑enabled with modern hardware (RTX6000 Pro gpu-latest runners) and more realistic training tests like coco128, thanks to PR #22710 by @glenn-jocher.
  • Clearer GPU auto‑selection logs and warnings when fewer idle GPUs than requested are available were added in PR #22716 by @glenn-jocher.
  • GitHub Actions runner images were bumped to v2.329.0 via PR #22707 by @glenn-jocher for a more modern CI base.

You should see more predictable behavior when using multiple GPUs, especially on small custom datasets.

:books: Documentation & Guides

This release includes a wide‑ranging docs refresh focused on modern hardware and practical deployments:

  • A large documentation review by @glenn-jocher in PR #22698 brings:
    • Consistent YOLO11 branding and updated licensing text.
    • Updated setup and deployment guides for Docker, AzureML, Raspberry Pi, DeepStream, Coral Edge TPU, Vertex AI, and more.
    • Refreshed dataset pages (COCO variants, MNIST, DOTA, pose & segment subsets, tracking, Explorer) with clearer structures and examples.
  • @lakshanthad updated the NVIDIA Jetson guide for Jetson AGX Thor (JetPack 7) including specs, flashing steps, and new benchmarks in PR #22699.
  • @RizwanMunawar embedded a step‑by‑step YOLO11 training video on the KITTI dataset page via PR #22685 to help new users get started visually.

All of this aims to make it easier to move from “quick experiment” to production‑ready YOLO on real devices.

:puzzle_piece: Docs Build, UI & Chat Enhancements

The docs experience and LLM chat helper were improved in several steps:

  • HTML minification was made more efficient and safer to avoid breaking scripts or code blocks in PR #22676 and PR #22678 by @glenn-jocher, preserving script, pre, code, and textarea content and improving cleanup.
  • The local docs server now has clearer logging and better error handling for a smoother local docs dev workflow, also in PR #22676.
  • The in‑docs chat widget was iteratively upgraded:
  • Several fixes around extra JS, Explorer page 404s, and quoting bugs in the chat were addressed in:

The result is a faster, more robust docs site with richer, context‑aware assistance.

:artist_palette: Repo & Config Quality

Code style and config consistency were also improved:

  • Prettier’s default print-width was updated to 120 and applied across JS/HTML/CSS and dataset YAML files in PR #22706 by @glenn-jocher, resulting in cleaner diffs and improved readability.
  • Many docs tables and argument macros for train, val, export, predict, solutions, and visualization were tightened up to better match current APIs and best practices.

:busts_in_silhouette: PR Highlights & Contributors

A big thank you to everyone who contributed to this release (in alphabetical order of PR title):

You can see all changes in the full v8.3.229 changelog diff.

:test_tube: How to Upgrade and Try It

Upgrade to v8.3.229 with:

pip install -U ultralytics

Then run a quick validation (for example on a YOLO11 segmentation model):

yolo val model=yolo11x-seg.pt data=coco.yaml split=val

You should notice much faster COCO segmentation validation compared to earlier versions, especially on larger val sets.

:speech_balloon: Feedback & Discussion

We’d love to hear how v8.3.229 works for you:

  • Do you see the expected speedup in COCO segmentation validation?
  • Are the new environment guards and dependency bounds helping with installs and exports?
  • Is the updated Jetson / dataset / deployment documentation clearer for your workflows?

Please share your benchmarks, issues, and suggestions in this Discourse thread or open an issue or discussion in the Ultralytics GitHub repository.

Thanks again to the entire YOLO community and Ultralytics team for driving these improvements forward. Your real‑world feedback is what keeps YOLO fast, stable, and practical. :rocket: