Ultralytics v8.4.183: Broader Hardware Support, More Reliable Workflows
Ultralytics YOLO v8.4.183 expands YOLO26 deployment on Axelera and Intel hardware, boosts batched inference throughput, and improves reliability across training, prediction, and evaluation. This release also brings fixes for dataset handling, exports, and Ultralytics Platform workflows.
Hardware and Deployment Highlights
- Every YOLO26 task on Axelera Metis: PR #26782 by @a-tangfan adds Voyager SDK 1.9.0 support for detection, instance segmentation, semantic segmentation, depth estimation, pose, OBB, and classification through export, prediction, and validation. Detection uses the NMS-free path by default.
- The documented COCO128 benchmark reaches 109 images/second at batch 32, compared with 40 images/second on SDK 1.8. Results depend on hardware and workload; batch 1 remains useful for latency-sensitive applications.
- Existing SDK 1.8 exports remain compatible. The updated setup requires
metis-dkms1.7.0 or newer and PyTorch below 2.15.
- Intel acceleration for LiteRT: PR #26779 by @onuralpszr adds Intel CPU and NPU inference support, including YOLO26 task types.
Training and Resume Improvements
- More reliable warmup and gradient accumulation: PR #26753 by @cainiao33 clamps the warmup accumulation ramp to its capped nominal value.
- Reduced MPS memory growth: PR #26744 by @vtarasv buckets target padding to prevent per-shape driver memory growth during training.
- Consistent Platform resumes: PR #26689 by @raimbekovm keeps resumed local training under the same Platform model.
- More robust checkpoint and logging resumes: PR #26777 by @raimbekovm fixes distillation resumes under restricted checkpoint loading, while PR #26772 by @raimbekovm restores W&B resumes on Windows without the
latest-runsymlink. - Better depth calibration selection: PR #26778 by @Y-T-G preserves an existing calibration when it scores best.
Data and Evaluation Fixes
- Correct ImageNet-subset validation: PR #26760 by @Nikhi00718 aligns subset folders with classifier class IDs. In the reported ImageNette test, this corrects validation accuracy from 9.45% to 70.37%—an evaluation fix, not a model-training gain.
- Improved UTF-8 handling: PR #26745 by @MohammadHijjawi97 preserves non-ASCII grounding annotations, and PR #26761 by @mottopanikeiku accepts labels containing a UTF-8 byte-order mark.
- COCO RLE segmentation conversion: PR #26774 by @Vaishnavi220506 adds mask decoding in
convert_coco. - Correct YOLOE validation setup: PR #26770 by @raimbekovm converts NDJSON validation data before prompt setup.
- More flexible source handling: PR #26781 by @raimbekovm makes URL-scheme and archive-extension comparisons case-insensitive.
Prediction, Export, and Visualization Fixes
- Consistent image inputs: PR #26741 by @Nikhi00718 scales 16-bit NumPy inputs to 8 bits, matching image-file prediction behavior.
- Reliable FastSAM exclusions: PR #26759 by @Nikhi00718 ensures negative points exclude masks regardless of prompt order.
- Safer INT8 calibration: PR #26750 by @Y-T-G fixes crashes with
rect=True, while PR #26783 by @Nikhi00718 preserves rectangular classification calibration shapes. - Grayscale LiteRT compatibility: PR #26747 by @aswanth-07 corrects layout detection for grayscale exports.
- Thin-image and depth-map handling: PR #26771 by @aswanth-07 keeps letterbox dimensions positive, and PR #26754 by @cainiao33 preserves two-dimensional depth maps for single-row, single-column, and 1 × 1 outputs.
- Correct semantic overlay colors: PR #26755 by @cainiao33 fixes swapped red and blue channels in training and validation mosaics.
- Additional reliability fixes: PR #26766 by @cainiao33 addresses warmup endpoints,
Conv2fusion, BOM source lists, image-opening errors, semantic mask names, and float64 OBB corners. - Safer temporary imports: PR #26756 by @Vaishnavi220506 restores pre-existing temporary module aliases.
Platform, Documentation, and Maintenance
- PR #26765 by @glenn-jocher documents Gemma 4 and Nemotron as hosted Platform annotation models.
- PR #26775 by @raimbekovm clarifies Platform visibility, Annotate similar, TensorRT JetPack 6 troubleshooting, and DOTA submission flags.
- PR #26785 by @glenn-jocher requires
ultralytics-thop>=2.2.3for functional linear and attention operation counts. - PR #26776 by @UltralyticsAssistant updates the self-hosted runner cleanup action.
- PR #26788 by @glenn-jocher packages the v8.4.183 release.
Thank You, Contributors!
Thanks to the YOLO community and Ultralytics team for these improvements. A special welcome to first-time contributors @vtarasv and @mottopanikeiku!
Try It and Share Your Feedback
Upgrade to the announced version:
pip install --upgrade ultralytics==8.4.183
Explore the v8.4.183 release notes for details, or review the full changelog to see everything included.
How does v8.4.183 perform in your workflow? Share your hardware, task, benchmarks, and any issues below—your feedback helps improve future releases.