Ultralytics v8.4.92 Released
Ultralytics v8.4.92 improves RKNN INT8 multi-batch export reliability, strengthens compiled inference and dataset validation, preserves segmentation boxes near image boundaries, and expands YOLO26 web deployment documentation. 
Review the v8.4.92 release on GitHub for the complete release details.
New Features and Documentation
YOLO26 deployment with LiteRT.js
Browser and Node.js deployment documentation now covers the official @ultralytics/yolo NPM package, including WebGPU acceleration, CPU/WASM fallback, and support for all six YOLO26 tasks.
Install the packages with:
npm i @ultralytics/yolo @litertjs/core
This update was contributed in PR #25087 by @onuralpszr.
Refreshed pose dataset documentation
Documentation for COCO-Pose, COCO8-Pose, Dog-Pose, Hand Keypoints, and Tiger-Pose now includes more accurate dataset sizes, splits, keypoint details, and download requirements. It also provides clearer guidance for annotating and training custom pose datasets with Ultralytics Platform.
COCO-Pose users should plan for approximately 27 GB for the complete initial download.
These improvements arrived in PR #25062 from @raimbekovm.
Bug Fixes
Reliable RKNN INT8 multi-batch export
RKNN calibration now uses a batch-1 ONNX graph while passing the requested inference batch through rknn_batch_size. This resolves failures with configurations such as batch=8 and adds regression coverage for calibration and runtime batch handling.
See PR #25094 by @glenn-jocher for the implementation.
Safer compiled model reuse
Predictor initialization now explicitly checks whether the model is None, preventing torch.compile wrappers from triggering unsupported truth-value checks during repeated inference.
The fix is included in PR #25092 by @glenn-jocher.
Correct validation split handling
Dataset validation now checks the split requested by the user, such as val or test, rather than always checking val. Missing splits produce a clear error instead of silently falling back to another split.
This behavior was corrected in PR #25093 by @glenn-jocher.
Better polygon-to-box conversion
segment2box() now preserves the visible portion of polygons crossing image boundaries. This reduces shrunken or collapsed boxes after perspective and related augmentations.
The update was contributed through PR #25086 by @JESUSROYETH.
Workflow Improvements
The contributor license workflow now uses the shared ultralytics/actions workflow, reducing duplicated configuration and limiting workflow permissions.
This change was made in PR #25084 by @glenn-jocher.
Get v8.4.92
Upgrade to the latest release with:
pip install -U ultralytics
You can inspect every change in the v8.4.91 to v8.4.92 comparison before updating.
Give v8.4.92 a try and let us know how the improved RKNN export, validation, augmentation, compiled inference, and YOLO26 web deployment workflows perform in your projects. Thanks to all contributors and the wider YOLO community for helping make this release possible! 