New Release: Ultralytics v8.3.232

Discussion
, ,
1

:rocket: Ultralytics v8.3.232 Release – OBB Fixes, Better Docs, Smoother Deployments

Ultralytics v8.3.232 is out! This is a focused stability and UX release that delivers a critical OBB export fix for production deployments, modernizes our ARM64 Docker base, and significantly improves documentation navigation, search, and SAM 3 coverage. :sparkles:

If you are exporting OBB models or rely on our docs/search for daily work, this update is highly recommended.

:glowing_star: Summary

  • :white_check_mark: OBB export behavior fixed for more reliable post-export NMS and box counts
  • :penguin: ARM64 Docker image updated to Ubuntu 24.04 for better security and compatibility
  • :wrench: New GPU/cuDNN compatibility helper added to docs for easier CUDA troubleshooting
  • :books: SAM 3 docs aligned with Meta’s release, clearly marking integration as in-progress
  • :magnifying_glass_tilted_left: Docs search upgraded with a chat-powered, keyboard-driven search experience
  • :compass: Docs navigation, links, and UI cleaned up, plus Mermaid diagram support

You can review the full release details in the Ultralytics v8.3.232 notes on the Ultralytics GitHub releases page.

:new_button: New Features

:abacus: GPU/cuDNN Compatibility Helper

A ready-to-run snippet was added to yolo-common-issues.md to quickly check:

  • GPU compute capability
  • cuDNN version
  • Whether the combination is likely compatible (especially important for older GPUs)

This comes from “Add GPU/cuDNN compatibility check to yolo-common-issues.md” by @akdanz.

Example usage pattern (simplified):

import torch

print("CUDA available:", torch.cuda.is_available())
if torch.cuda.is_available():
    print("GPU name:", torch.cuda.get_device_name(0))
    print("CUDA capability:", torch.cuda.get_device_capability(0))
    print("cuDNN version:", torch.backends.cudnn.version())

This should cut down on guesswork when debugging CUDA/cuDNN issues across different stacks.

:wrench: Core & Deployment Improvements

:construction: OBB Export Fix for More Reliable Deployments

OBB export behavior for detection models has been corrected so exported OBB models (including quantized formats like TFLite int8) behave more consistently with pre-export behavior.

Key changes:

  • Uses a fixed multiplier of 8 for OBB NMS in ExporterDetectionModel.forward
  • Non-OBB models keep the existing logic (normalized by number of classes)

This prevents under-detection and overly-aggressive NMS in exported OBB models. The change was contributed in “ultralytics 8.3.232 Fix OBB NMS export lower box count” by @Y-T-G.

:penguin: ARM64 Docker Image Modernization

For ARM64 users (Apple Silicon, Raspberry Pi, and other ARM boards), the Docker base has been updated for better security and compatibility:

  • Base switched from arm64v8/debian:bookworm-slim to arm64v8/ubuntu:24.04
  • Uses apt with apt upgrade -y to keep system packages up-to-date

This work comes from “fix: :lady_beetle: update Dockerfile to use Ubuntu 24.04 LTS and streamline package installation commands” by @onuralpszr.

:books: Documentation & UX Upgrades

:blue_book: SAM 3 Documentation Aligned with Meta’s Release

Docs now clearly reflect that Meta has officially released SAM-3, and that Ultralytics integration is in progress. Usage will remain via pip install ultralytics, and all SAM 3 content is clearly marked as API previews until integration lands.

These updates were made in “SAM3 docs update” by @glenn-jocher.

:magnifying_glass_tilted_left: Chat-Powered Docs Search

The documentation search experience received a major upgrade:

  • A custom chat-based search button was added to the docs header with OS-aware hotkeys (⌘K / Ctrl+K)
  • The default MkDocs search UI is hidden in favor of this richer experience
  • Theme-aware CSS variables ensure proper light/dark mode styling
  • Mobile layout is optimized with an icon-only, touch-friendly button

This is implemented across multiple PRs:

:compass: Docs Navigation, Links & Build Tweaks

Several quality-of-life improvements make the docs easier to navigate and maintain:

  • TOC .md links are now standardized and rewritten as trailing-slash internal paths
  • Internal links on the YOLO history page and homepage use consistent paths (for example, internal links like models/yolov5.md are normalized)
  • Navigation labels are cleaned up (for example, “Modes:”, “Tasks:”, “Models:”, “HUB:”)
  • Docs build command no longer uses --strict in the Zensical build, reducing failures caused by minor warnings
  • The Weglot language switcher is hidden via CSS for a cleaner UI
  • New contributors are added to mkdocs_github_authors.yaml for proper attribution

These updates were delivered across several PRs:

:bar_chart: Mermaid Diagram Support in Docs

Mermaid support was enabled via pymdownx.superfences configuration in mkdocs.yml, allowing rich diagrams such as flowcharts and sequence diagrams directly in docs using fenced ```mermaid blocks.

This feature is added in “Docs Mermaid diagram support” by @glenn-jocher.

:raising_hands: Contributors & Community

This release is a great example of the community and Ultralytics team working together to polish the ecosystem.

You can see the complete list of changes in the full changelog comparing v8.3.231…v8.3.232.

:bullseye: Why This Matters

  • Production users of OBB models get more consistent, reliable results after export, especially under int8 quantization.
  • ARM64 users benefit from a modern Ubuntu 24.04 LTS base, improving compatibility and security.
  • Anyone debugging CUDA issues can quickly assess GPU/cuDNN compatibility using the new helper snippet.
  • All users of the docs get a smoother navigation experience, better search, and richer visual explanations via Mermaid diagrams.

Overall, v8.3.232 is about stability, clarity, and deployment confidence, especially around OBB exports and documentation.

:rocket: Try It Out & Share Feedback

Update to the latest release with:

pip install -U ultralytics

Then run, for example:

yolo predict model=yolo11n.pt source=path/to/your/images

Feedback, ideas, and bug reports are always welcome in the Ultralytics GitHub discussions and issues. The community’s input is what keeps YOLO improving for everyone.

If you try v8.3.232 in your OBB pipelines, ARM64 environments, or rely heavily on the docs, please share:

  • Any changes you see in exported OBB performance
  • How the new search and navigation feel in your day-to-day use
  • Suggestions for further tooling or docs improvements

Thank you to everyone contributing, testing, and helping push the ecosystem forward. :light_bulb: