New Release: Ultralytics v8.4.32

Discussion
, ,
1

Ultralytics v8.4.32 is out :tada:

Quick summary: v8.4.32 is a deployment-focused release centered on a major Axelera AI export expansion :rocket:, with cleaner export architecture, broader task support across Ultralytics YOLO models, new end-to-end examples, and several helpful documentation improvements for both the repo and the Ultralytics Platform.

If you deploy on edge hardware, especially Axelera devices, this is a great release to try. If you use Platform for dataset workflows, this update also makes a few important behaviors and limits much clearer :white_check_mark:

:glowing_star: Highlights

Axelera export expansion and refactor

The biggest change in v8.4.32 comes from PR #23844 by @lakshanthad, which significantly improves Axelera export and deployment workflows.

Key upgrades include:

  • Moved Axelera export logic into a dedicated utility module: ultralytics/utils/export/axelera.py
  • Introduced reusable torch2axelera helper for a cleaner export pipeline
  • Expanded export support beyond detection-only messaging to better cover multiple YOLO task workflows across YOLOv8, YOLO11, and YOLO26
  • Added task-aware calibration defaults via TASK2CALIBRATIONDATA
  • Clarified export requirements and limitations, including:
    • Linux-only support
    • no ARM64 Docker support
    • Torch 2.8+ requirement

This makes the Axelera path easier to maintain, easier to understand, and more practical for real deployment scenarios :brain::gear:

Better Axelera runtime support

This release also updates Axelera backend/runtime handling with newer package paths and more direct .axm pipeline loading and optimization, helping smooth the path from export to inference :electric_plug:

New deployment examples

v8.4.32 adds practical Python deployment examples for:

  • YOLO26 pose + tracking
  • YOLO11 segmentation

These examples are especially useful because they demonstrate full inference pipelines on Axelera hardware without requiring Ultralytics at inference runtime :file_folder::bullseye:

:books: Documentation improvements

A big part of this release is improved documentation across both the Ultralytics repo and the Ultralytics Platform docs.

Axelera docs updates

The Axelera integration docs were expanded with:

  • supported tasks tables
  • Python version support for 3.10, 3.11, and 3.12
  • clearer installation steps
  • known limitations
  • benchmark tables
  • new reference docs for the Axelera export utility

Platform docs clarifications

This release also improves Platform documentation to reduce confusion in day-to-day workflows:

Docs UI polish

A few nice usability improvements landed too:

:wrench: Why this release matters

For developers

  • Cleaner export internals mean easier maintenance and faster iteration
  • Reusable Axelera utilities should reduce regression risk
  • Better calibration defaults mean less manual export setup

For edge deployment users

  • Broader support across YOLOv8, YOLO11, and YOLO26 improves real-world deployment flexibility
  • New examples help bridge the gap from trained model to production-style inference
  • More explicit limitations make setup more predictable

For Platform users

  • Clearer dataset upload format and size documentation reduces workflow friction
  • Better annotation behavior documentation helps avoid accidental assumptions
  • Small UI fixes improve the docs experience overall

:receipt: What changed

For the full diff, check the full changelog from v8.4.31 to v8.4.32, and you can browse the complete release on the Ultralytics v8.4.32 release page.

:raising_hands: Try it out

We’d love for the community to test v8.4.32, especially if you’re working with Axelera deployments, multi-task annotation workflows, or the latest Ultralytics YOLO models like YOLO26.

If you give it a spin, please share your feedback, results, and any edge cases you run into. Your input helps make each release better for everyone :blue_heart:

1 Like