Ultralytics v8.4.24 is out 
Quick summary: Ultralytics v8.4.24 improves training reliability on the Ultralytics Platform
, updates tuning defaults to better match Ultralytics YOLO26 best practices
, refreshes docs and benchmarks, and adds safer guardrails around TF.js benchmarking.
This release is all about making training workflows clearer, tuning more effective, and the overall experience more consistent across product, docs, and tooling.
Highlights
Better Platform training error surfacing 
The biggest improvement in v8.4.24 is clearer error handling for Platform training sessions.
Users will now see the actual server-side error message when available, instead of a generic HTTP response. If a training session fails to register, the message is now much clearer:
Training will not be tracked on Platform
To keep logs clean and avoid repeated follow-up warnings, Platform callbacks are also disabled after registration failure.
This means:
- Faster troubleshooting

- Less noisy logs

- A smoother fallback experience

Ray Tune search space updated for YOLO26 
Weβve aligned the Ray Tune search space and examples with modern YOLO26 tuning ranges, including updates to parameters like:
lr0momentumboxclsscale
We also added missing parameters such as:
dflclose_mosaic
These updates help users tune with more realistic defaults and avoid outdated search spaces that can hurt results.
YOLO26 naming adopted in the Streamlit selector 
The Streamlit inference UI now uses yolo26* model options instead of yolo11*, bringing the experience in line with the latest recommended model family.
Rockchip RKNN docs refreshed 
RKNN benchmark tables were updated to reflect YOLO26 models and newer tested package versions, making deployment guidance more current and useful.
TF.js benchmarking temporarily disabled 
To prevent confusing runtime failures caused by a known protobuf dependency conflict, TF.js export benchmarking has been temporarily blocked as a safeguard.
Improvements
Docs and UX polish 
This release also includes several documentation and interface improvements:
- Clearer Platform plan and GPU tier messaging
- Added real screenshots to important Platform docs pages
- Formatting and content cleanup across docs
- Small UI and reference page polish
What changed
- Fix Platform cloud deployments docs by @glenn-jocher
- Missing screenshots placeholder updates by @t-hakobyan
- Add
UTMenabled banner redirect by @RizwanMunawar - Update Rockchip RKNN benchmarks with YOLO26 by @lakshanthad
- Migrate model prefixes from
YOLO11toYOLO26by @RizwanMunawar - Remove duplicate Dataset Tabs heading by @amanharshx
- Sync Ray Tune search space and docs with native tuner ranges by @raimbekovm
- Scope class badges to reference section by @glenn-jocher
ultralytics 8.4.24Improve Platform train error surfacing by @glenn-jocher
New contributors
A big welcome to our new contributors:
- @t-hakobyan for their first contribution in PR #23932
- @amanharshx for their first contribution in PR #23940
Try it out
Upgrade with:
pip install -U ultralytics
Then explore the latest release in the v8.4.24 release notes, or review everything included in the full changelog.
If youβre building new workflows, we recommend starting with Ultralytics YOLO26 and the Ultralytics Platform documentation for the most up-to-date training and deployment experience.
Feedback welcome 
Please give v8.4.24 a try and let us know how it goes. Feedback, bug reports, and PRs from the community continue to make YOLO better for everyone.