NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
PyPI
PyPI, #1876 most downloaded on PyPI
Ultralytics YOLO 🚀 for SOTA object detection, instance segmentation, semantic segmentation, depth estimation, classification, pose estimation, oriented object detection, and multi-object tracking.
Latest stable
8.4.210
Last release 11 days ago
10 Nov 2026
Ships on a steady schedule
a new release about every 8 days
Nearly every release is documented
notes for 60 of the last 60 stable releases
11 versions withdrawn
withdrawn after publishing
5 years old
885 releases · first in 2022
Drag across the ruler to compare, or tap a release to open it.
The intel:cpu, intel:gpu, and intel:npu names still work as deprecated aliases, with a warning. Device names are consistent, but their hardware mappin…
Release v8.4.210 makes hardware selection more consistent across model formats, improves OCR for rotated text, and fixes several deployment and inference issues.
cpu for Intel CPU, xpu for Intel GPU, and npu for an Intel NPU with those runtimes. The intel:cpu, intel:gpu, and intel:npu names still work as deprecated aliases, with a warning. Device names are consistent, but their hardware mapping depends on the runtime—for example, PyTorch npu refers to Ascend, while OpenVINO and LiteRT use Intel NPU.cpu selection also now reliably means CPU rather than automatic device selection.Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.209...v8.4.210
Your coding agent can read these notes before it upgrades. Set up the MCP server →
v8.4.210 - ultralytics 8.4.210 Use PyTorch device names for OpenVINO and LiteRT Intel CPU, GPU and NPU (#27136)
Compare
v8.4.209 makes concurrent YOLO predictions more reliable and includes targeted fixes for validation, exports, and device-specific inference.
v8.4.209 makes concurrent YOLO predictions more reliable and includes targeted fixes for validation, exports, and device-specific inference.
val(classes=[...]) now filters predictions as well as ground-truth labels, so excluded classes are less likely to affect metrics or saved results.batch > 1 are now rejected before conversion, rather than potentially producing a model that crashes during prediction.YOLO("rtdetr-l.yaml") builds by @cainiao33 in https://github.com/ultralytics/ultralytics/pull/27099Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.208...v8.4.209
v8.4.209 - Correct results for every request when serving YOLO to many users at once (#27140)
Compare
Ultralytics 8.4.208 ships practical inference, visualization, and performance updates—most notably a 0.5 default NMS IoU for prediction and tracking ,
Ultralytics 8.4.208 ships practical inference, visualization, and performance updates—most notably a 0.5 default NMS IoU for prediction and tracking, while validation keeps its 0.7 default.
iou is not specified, prediction and tracking now use 0.5 instead of 0.7. Validation remains at 0.7, and explicitly provided values still take precedence.conf=0.05, helping retain more similar objects whose confidence scores may be low. Other prediction calls are unchanged.iou when needed.ultralytics 8.4.208 Release the 0.5 predict NMS IoU default by @glenn-jocher in #27102Full Changelog: v8.4.207...v8.4.208
v8.4.208 - ultralytics 8.4.208 Release the 0.5 predict NMS IoU default (#27102)
Compare
Ultralytics 8.4.207 restores reliable PaddlePaddle export and inference, while refreshing several Docker base images.
Ultralytics 8.4.207 restores reliable PaddlePaddle export and inference, while refreshing several Docker base images.
six dependency required by x2paddle, which could break exports in fresh environments.ultralytics 8.4.207 Fix PaddlePaddle export on new installs and CUDA machines by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/27088Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.206...v8.4.207
v8.4.207 - ultralytics 8.4.207 Fix PaddlePaddle export on new installs and CUDA machines (#27088)
Compare
Ultralytics 8.4.206 improves PaddleOCR’s handling of upside-down text and makes OCR/rotated-box labels easier to read. 🔄
Ultralytics 8.4.206 improves PaddleOCR’s handling of upside-down text and makes OCR/rotated-box labels easier to read. 🔄
paddleocr>=3.7 and PaddlePaddle 3.1 or later. Custom recognizers are unchanged and do not get automatic orientation correction.macos-latest and ubuntu-26.04-arm labels, reducing future maintenance as runner images change.ultralytics 8.4.206 Read upside-down and bottom-to-top text in OCR by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/27083Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.205...v8.4.206
v8.4.206 - ultralytics 8.4.206 Read upside-down and bottom-to-top text in OCR (#27083)
Compare
Release 8.4.205 adds a ready-to-use scene-text OCR solution, led by YOLO26-Text detection paired with PaddleOCR or a custom recognizer, alongside impo
Release 8.4.205 adds a ready-to-use scene-text OCR solution, led by YOLO26-Text detection paired with PaddleOCR or a custom recognizer, alongside important tiled-inference and export fixes.
Results API remains unchanged.yolo checks now ignores requirements that do not apply to the current environment, and dataset split fractions reject booleans instead of treating them as numbers.yolo checks by @MohammadHijjawi97 in https://github.com/ultralytics/ultralytics/pull/27078ultralytics 8.4.205 Add scene text OCR solution with YOLO26-Text and PaddleOCR by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/27061Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.204...v8.4.205
v8.4.205 - ultralytics 8.4.205 Add scene text OCR solution with YOLO26-Text and PaddleOCR (#27061)
Compare
The SAHI guide remains available with a deprecation notice directing users to the built-in tile option, and the video example now uses the standard YO…
Ultralytics 8.4.204 makes high-resolution tiling available across training, validation, and prediction, and adds improvements to FastSAM prompts, model exports, and dataset handling.
tile=True or a positive integer tile size to process large images in overlapping windows. Training crops labels to each tile; validation and prediction merge results back into the original image.tile option, and the video example now uses the standard YOLO CLI and Python API.Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.203...v8.4.204
v8.4.204 - Release 8.4.204 with High Resolution Tiling documentation (#27069)
Compare
v8.4.203 adds NXP i.MX95 Neutron NPU export for YOLO models, plus security improvements and fixes across training, inference, tracking, and visualizat
v8.4.203 adds NXP i.MX95 Neutron NPU export for YOLO models, plus security improvements and fixes across training, inference, tracking, and visualization.
format=nxp to create an INT8 LiteRT model compiled using Neutron SDK 3.3.0. The export supports all seven YOLO tasks and uses the existing LiteRT inference backend with NXP’s delegate. It targets i.MX95 only; unsupported operations may run on the CPU.Annotator.text with PIL fonts by @MohammadHijjawi97 in https://github.com/ultralytics/ultralytics/pull/27038Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.202...v8.4.203
v8.4.203 adds NXP i.MX95 Neutron NPU export for YOLO models, plus security improvements and fixes across training, inference, tracking, and visualization.
format=nxp to create an INT8 LiteRT model compiled using Neutron SDK 3.3.0. The export supports all seven YOLO tasks and uses the existing LiteRT inference backend with NXP’s delegate. It targets i.MX95 only; unsupported operations may run on the CPU.Annotator.text with PIL fonts by @MohammadHijjawi97 in #27038Full Changelog: v8.4.202...v8.4.203
v8.4.203 - Add NXP i.MX95 Neutron export (#27047)
Compare
Ultralytics 8.4.202 improves stream-video saving and fixes several compatibility and reliability issues, alongside updated model and benchmark documen
Ultralytics 8.4.202 improves stream-video saving and fixes several compatibility and reliability issues, alongside updated model and benchmark documentation.
.streams sources use each stream’s own FPS—including in multi-stream setups—instead of assuming 30 FPS. Video-file and screen capture rates remain unchanged.val nor test split exists. Platform training arguments are sent as JSON values rather than being converted to strings.ultralytics 8.4.202 Save stream prediction videos at the stream's frame rate by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/27016Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.201...v8.4.202
v8.4.202 - ultralytics 8.4.202 Save stream prediction videos at the stream's frame rate (#27016)
Compare
v8.4.201 makes PIL-based plotting up to 15× faster while keeping visuals consistent, and standardizes the Construction-PPE dataset’s name and download
v8.4.201 makes PIL-based plotting up to 15× faster while keeping visuals consistent, and standardizes the Construction-PPE dataset’s name and download paths. 🚀
Results.plot(pil=True) benefit from cached fonts and label patches, fewer image conversions, and more efficient mask blending.uint8 to reduce overhead.construction-ppe archive and directory. The dataset is described as an original Ultralytics dataset, created and open-sourced in November 2026, with 1,000 images, 12 classes, and a CC BY 4.0 license.ultralytics 8.4.201 Up to 15x faster PIL plotting across all tasks by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/27014Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.200...v8.4.201
v8.4.201 - ultralytics 8.4.201 Up to 15x faster PIL plotting across all tasks (#27014)
Compare
Ultralytics 8.4.200 displays Arabic class names correctly in plots and makes training, export, and dataset workflows more reliable. No model architect
Ultralytics 8.4.200 displays Arabic class names correctly in plots and makes training, export, and dataset workflows more reliable. No model architecture changes are included.
%YAML or %TAG headers can now be parsed instead of being rejected.ultralytics 8.4.200 Display Arabic class names correctly in plots by @DrEssam97 in https://github.com/ultralytics/ultralytics/pull/27007Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.199...v8.4.200
v8.4.200 - ultralytics 8.4.200 Display Arabic class names correctly in plots (#27007)
Compare
Ultralytics 8.4.199 improves how training runs appear on Ultralytics Platform, makes several training and profiling workflows more reliable, and clari
Ultralytics 8.4.199 improves how training runs appear on Ultralytics Platform, makes several training and profiling workflows more reliable, and clarifies Platform and dataset features in the documentation. No model architectures changed in this release.
labels.jpg now show the foreground class rather than the background for single-class semantic datasets.ultralytics 8.4.199 Send the raw run name to Ultralytics Platform by @laodouya in https://github.com/ultralytics/ultralytics/pull/26953Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.198...v8.4.199
v8.4.199 - ultralytics 8.4.199 Send the raw run name to Ultralytics Platform (#26953)
Compare
Ultralytics 8.4.198 makes training settings and YAML files more reliable, alongside fixes to downloads, dataset errors, and documentation. No model ar
Ultralytics 8.4.198 makes training settings and YAML files more reliable, alongside fixes to downloads, dataset errors, and documentation. No model architecture or pretrained-weight changes are included.
np.float64 values to ordinary numeric values when writing YAML. Training arguments such as sweep-generated learning rates can now be saved and reused from args.yaml.ultralytics-thop and ultralytics-platform; also suppresses a Triton warning that could appear during profiling.project paths are resolved and standardizes how metrics such as mAP and depth accuracy are written across docs and benchmark tables.project goes under runs_dir/<task> by @Echoslayer in https://github.com/ultralytics/ultralytics/pull/26985ultralytics 8.4.198 Save NumPy float64 values as numbers in YAML files by @Nikhi00718 in https://github.com/ultralytics/ultralytics/pull/26993Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.197...v8.4.198
Ultralytics 8.4.198 makes training settings and YAML files more reliable, alongside fixes to downloads, dataset errors, and documentation. No model architecture or pretrained-weight changes are included.
np.float64 values to ordinary numeric values when writing YAML. Training arguments such as sweep-generated learning rates can now be saved and reused from args.yaml.ultralytics-thop and ultralytics-platform; also suppresses a Triton warning that could appear during profiling.project paths are resolved and standardizes how metrics such as mAP and depth accuracy are written across docs and benchmark tables.project goes under runs_dir/<task> by @Echoslayer in #26985ultralytics 8.4.198 Save NumPy float64 values as numbers in YAML files by @Nikhi00718 in #26993Full Changelog: v8.4.197...v8.4.198
v8.4.198 - ultralytics 8.4.198 Save NumPy float64 values as numbers in YAML files (#26993)
Compare
Ultralytics 8.4.197 makes AMP checks work without downloading model weights and removes PyYAML from the base dependencies. 🚀
Ultralytics 8.4.197 makes AMP checks work without downloading model weights and removes PyYAML from the base dependencies. 🚀
ultralytics.utils.YAML to save YAML files.ultralytics 8.4.197 Check AMP on an untrained YOLO26n without downloading weights by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/26967Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.196...v8.4.197
v8.4.197 - ultralytics 8.4.197 Check AMP on an untrained YOLO26n without downloading weights (#26967)
Compare
Release v8.4.196 improves ONNX INT8 model accuracy and makes several practical fixes to training, video analytics, exports, and documentation.
Release v8.4.196 improves ONNX INT8 model accuracy and makes several practical fixes to training, video analytics, exports, and documentation.
ultralytics 8.4.196 Improve ONNX INT8 accuracy with per-channel weight quantization by @amanharshx in https://github.com/ultralytics/ultralytics/pull/26978Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.195...v8.4.196
v8.4.196 - ultralytics 8.4.196 Improve ONNX INT8 accuracy with per-channel weight quantization (#26978)
Compare
v8.4.195 adds early Python 3.15 compatibility checks, fixes AutoBatch profiling on older PyTorch versions, and improves documentation accuracy across
v8.4.195 adds early Python 3.15 compatibility checks, fixes AutoBatch profiling on older PyTorch versions, and improves documentation accuracy across hundreds of pages.
ultralytics 8.4.195 Add Python 3.15 CI and classifier by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/26975Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.194...v8.4.195
v8.4.195 - ultralytics 8.4.195 Add Python 3.15 CI and classifier (#26975)
Compare
Ultralytics 8.4.194 makes automatic batch sizing more reliable across tasks, alongside practical fixes for model loading, results saving, datasets, an
Ultralytics 8.4.194 makes automatic batch sizing more reliable across tasks, alongside practical fixes for model loading, results saving, datasets, and deployment workflows.
torch.nn modules, and model entries must be PyTorch module classes. Custom YAMLs using other expressions may need updating.Results.save() and plot(save=True) now produce sensible .jpg filenames for sources such as webcams, videos, or missing paths, and create parent directories when needed.HTTPS://, GS://, and UL:// are now handled without case sensitivity.Results.save and plot default filename when path is None or extensionless by @Rahulbiradar9 in https://github.com/ultralytics/ultralytics/pull/26958parse_model by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/26963ultralytics 8.4.194 Size AutoBatch from the train dataset for every task by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/26956Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.193...v8.4.194
v8.4.194 - ultralytics 8.4.194 Size AutoBatch from the train dataset for every task (#26956)
Compare
Ultralytics 8.4.193 makes CPU inference with ONNX Runtime and OpenVINO substantially faster on some systems, alongside fixes that improve export behav
Ultralytics 8.4.193 makes CPU inference with ONNX Runtime and OpenVINO substantially faster on some systems, alongside fixes that improve export behavior, data handling, training reliability, and documentation.
.pt setup is unchanged.imgsz, prediction and validation now default to the size recorded in the export metadata. This avoids silently running models at an unsuitable default resolution; an explicitly provided size still takes precedence.uint8 labels correctly; and W&B class curves now skip label-less validation splits and use the right class names. YOLOE now clearly rejects visual prompts on fused models rather than failing later or reporting misleading results.uv when available—including on ARM64. The release also updates CI coverage for newer PyTorch versions on Python 3.14..pt setup.imgsz is omitted.ultralytics 8.4.193 Up to 2x faster CPU ONNX Runtime and OpenVINO prediction by @var1914 in https://github.com/ultralytics/ultralytics/pull/26942Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.192...v8.4.193
Ultralytics 8.4.193 makes CPU inference with ONNX Runtime and OpenVINO substantially faster on some systems, alongside fixes that improve export behavior, data handling, training reliability, and documentation.
.pt setup is unchanged.imgsz, prediction and validation now default to the size recorded in the export metadata. This avoids silently running models at an unsuitable default resolution; an explicitly provided size still takes precedence.uint8 labels correctly; and W&B class curves now skip label-less validation splits and use the right class names. YOLOE now clearly rejects visual prompts on fused models rather than failing later or reporting misleading results.uv when available—including on ARM64. The release also updates CI coverage for newer PyTorch versions on Python 3.14..pt setup.imgsz is omitted.ultralytics 8.4.193 Up to 2x faster CPU ONNX Runtime and OpenVINO prediction by @var1914 in #26942Full Changelog: v8.4.192...v8.4.193
v8.4.193 - ultralytics 8.4.193 Up to 2x faster CPU ONNX Runtime and OpenVINO prediction (#26942)
Compare
v8.4.192 improves compatibility with PyTorch 2.15 and CUDA 13 for multi-GPU training and ONNX Runtime GPU inference, alongside a small update to Platf
v8.4.192 improves compatibility with PyTorch 2.15 and CUDA 13 for multi-GPU training and ONNX Runtime GPU inference, alongside a small update to Platform response documentation.
metadata.classNames. Class names remain available on individual detections; dataset and model class names are unchanged.No model architecture or prediction behavior changes are indicated in this release.
Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.191...v8.4.192
v8.4.192 - Support torch 2.15 CUDA in DDP training and ONNX Runtime inference (#26928)
Compare
Release v8.4.191 improves RKNN INT8 exports, especially for tasks beyond detection, and includes performance and tracking fixes plus Platform document
Release v8.4.191 improves RKNN INT8 exports, especially for tasks beyond detection, and includes performance and tracking fixes plus Platform documentation updates.
tail.onnx file runs the box decoding on the CPU in floating point. FP16-capable targets also use RKNN’s auto_hybrid option to retain high-error layers in floating point.tail.onnx or implement the same decoding. Re-export INT8 models made with earlier versions to avoid incorrect coordinate scaling. INT8 accuracy has been tested in the RKNN simulator, but not yet validated on Rockchip hardware.delta_t-frame window, making associations less sensitive to box jitter.ultralytics 8.4.191 Fix RKNN INT8 accuracy with a host-side decode tail for all tasks by @Bovey0809 in https://github.com/ultralytics/ultralytics/pull/26904Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.190...v8.4.191
v8.4.191 - ultralytics 8.4.191 Fix RKNN INT8 accuracy with a host-side decode tail for all tasks (#26904)
Compare
YOLO 8.4.190 improves YOLO26 segmentation training by making label assignment consider mask quality as well as box overlap, alongside fixes for model
YOLO 8.4.190 improves YOLO26 segmentation training by making label assignment consider mask quality as well as box overlap, alongside fixes for model export, inference, training utilities, and data handling.
ultralytics 8.4.190 Mask-IoU-aware label assignment for YOLO26-seg by @Bovey0809 in https://github.com/ultralytics/ultralytics/pull/26773Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.189...v8.4.190
YOLO 8.4.190 improves YOLO26 segmentation training by making label assignment consider mask quality as well as box overlap, alongside fixes for model export, inference, training utilities, and data handling.
ultralytics 8.4.190 Mask-IoU-aware label assignment for YOLO26-seg by @Bovey0809 in #26773Full Changelog: v8.4.189...v8.4.190
v8.4.190 - ultralytics 8.4.190 Mask-IoU-aware label assignment for YOLO26-seg (#26773)
Compare
Ultralytics 8.4.189 improves deployment reliability and inference robustness, with key fixes for Sony IMX500 and Hailo exports, clearer dataset errors
Ultralytics 8.4.189 improves deployment reliability and inference robustness, with key fixes for Sony IMX500 and Hailo exports, clearer dataset errors, and safer handling of model outputs. No model architecture changes were introduced.
einsum operations with equivalent matrix multiplications, enabling affected YOLOE and YOLO-World models to export to ONNX at lower opsets.ultralytics 8.4.189 Fix IMX requirements, Hailo seg end nodes and unreadable-dataset errors by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/26883Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.188...v8.4.189
Ultralytics 8.4.189 improves deployment reliability and inference robustness, with key fixes for Sony IMX500 and Hailo exports, clearer dataset errors, and safer handling of model outputs. No model architecture changes were introduced.
einsum operations with equivalent matrix multiplications, enabling affected YOLOE and YOLO-World models to export to ONNX at lower opsets.ultralytics 8.4.189 Fix IMX requirements, Hailo seg end nodes and unreadable-dataset errors by @glenn-jocher in #26883Full Changelog: v8.4.188...v8.4.189
v8.4.189 - ultralytics 8.4.189 Fix IMX requirements, Hailo seg end nodes and unreadable-dataset errors (#26883)
Compare
Release 8.4.188 improves QNN exports for Android, removes an unnecessary operation from non-LiteRT depth exports, and updates the LiteRT export toolch
Release 8.4.188 improves QNN exports for Android, removes an unnecessary operation from non-LiteRT depth exports, and updates the LiteRT export toolchain.
io_name_overrides attribute that newer QNN export tools can add but Android’s built-in QNN Execution Provider cannot read.litert-torch 0.10.0, ai-edge-litert 2.3.0, and ai-edge-quantizer 0.10.0.Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.187...v8.4.188
v8.4.188 - ultralytics 8.4.188 Fix Android loading and depth speed of QNN exports (#26872)
Compare
YOLO26-pose training now assigns keypoints using pose accuracy, with a feature-pyramid scale gate to help match objects to suitable training anchors.
YOLO26-pose training now assigns keypoints using pose accuracy, with a feature-pyramid scale gate to help match objects to suitable training anchors.
Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.186...v8.4.187
v8.4.187 - ultralytics 8.4.187 OKS-only label assignment with a pyramid-level gate for YOLO26-pose (#26843)
Compare
Improves compatibility with newer Linux systems by fixing imports for some older PyTorch versions and MNN on glibc 2.41 and later.
Improves compatibility with newer Linux systems by fixing imports for some older PyTorch versions and MNN on glibc 2.41 and later.
import ultralytics and MNN imports work on affected x86_64 Linux systems, including Ubuntu 26.04.Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.185...v8.4.186
v8.4.186 - Load torch 1.10-1.12 on glibc 2.41+ with one shared executable-stack fix for torch and MNN (#26866)
Compare
This release primarily fixes MNN loading on newer Linux systems, while improving ONNX deployment options, export reliability, and several training and
This release primarily fixes MNN loading on newer Linux systems, while improving ONNX deployment options, export reliability, and several training and inference workflows.
onnx2tf 1.x version across Python versions. In reported testing, this changed a YOLO26 export from 20 to 652 operations mapped to the Edge TPU, with far fewer left for the CPU.onnx2tf adjustment helps more exported model operations run on Edge TPU rather than falling back to CPU.Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.184...v8.4.185
v8.4.185 - Load MNN on glibc 2.41+ by clearing its unused executable-stack flag (#26861)
Compare
A reliability-focused release with an important TIFF-reading fix, plus improvements to segmentation prompts, model loading, validation, integrations,
A reliability-focused release with an important TIFF-reading fix, plus improvements to segmentation prompts, model loading, validation, integrations, and Platform workflows—without introducing a new model architecture.
cv2.IMREAD_UNCHANGED stacks same-size pages, preserving the option for multispectral workflows.best.pt. Changing class names on a text-fused YOLOE model now raises a clear error instead of silently relabeling predictions or failing later.rsync, and libraries needed to import GUI OpenCV. Documentation clarifies Platform dataset workflows and training features, corrects export examples, and refreshes several guides and screenshots.cv2.IMREAD_UNCHANGED.Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.183...v8.4.184
A reliability-focused release with an important TIFF-reading fix, plus improvements to segmentation prompts, model loading, validation, integrations, and Platform workflows—without introducing a new model architecture.
cv2.IMREAD_UNCHANGED stacks same-size pages, preserving the option for multispectral workflows.best.pt. Changing class names on a text-fused YOLOE model now raises a clear error instead of silently relabeling predictions or failing later.rsync, and libraries needed to import GUI OpenCV. Documentation clarifies Platform dataset workflows and training features, corrects export examples, and refreshes several guides and screenshots.cv2.IMREAD_UNCHANGED.Full Changelog: v8.4.183...v8.4.184
v8.4.184 - Read multi-page grayscale TIFFs as their first page unless every band is requested (#26842)
Compare
Ultralytics 8.4.183 expands hardware deployment and improves reliability across training, prediction, data handling, and evaluation. 🚀
Ultralytics 8.4.183 expands hardware deployment and improves reliability across training, prediction, data handling, and evaluation. 🚀
yolo. The NMS-free detection path is used by default. SDK 1.8 exports remain compatible; the updated setup requires metis-dkms 1.7.0 or newer and supports PyTorch versions below 2.15.Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.182...v8.4.183
Ultralytics 8.4.183 expands hardware deployment and improves reliability across training, prediction, data handling, and evaluation. 🚀
yolo. The NMS-free detection path is used by default. SDK 1.8 exports remain compatible; the updated setup requires metis-dkms 1.7.0 or newer and supports PyTorch versions below 2.15.Full Changelog: v8.4.182...v8.4.183
v8.4.183 - ultralytics 8.4.183 Update Axelera SDK to 1.9.0 and support every YOLO26 task (#26788)
Compare
YOLO 8.4.182 strengthens dataset label checks to catch invalid bounding boxes before they break validation plotting, and improves Linux drive reportin
YOLO 8.4.182 strengthens dataset label checks to catch invalid bounding boxes before they break validation plotting, and improves Linux drive reporting.
SystemLogger drive discovery by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/26746Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.181...v8.4.182
v8.4.182 - Reject negative box dimensions during label verification (#26748)
Compare
Version 8.4.181 focuses on making dataset loading, annotation, tracking, and inference more reliable, with several performance and compatibility impro
Version 8.4.181 focuses on making dataset loading, annotation, tracking, and inference more reliable, with several performance and compatibility improvements. No model architecture changes are included.
.yaml files before .yml files, preventing unrelated YAML files from breaking dataset loading. .txt and .csv prediction-source lists work regardless of filename capitalization, and dataset hashing now supports filenames with non-UTF-8 characters.auto_annotate now names label files using the detector’s prediction output path, helping keep labels matched to images with the same filename stem..yml dataset discovery regression, case-insensitive source lists, non-UTF-8 hashing and auto_annotate same-stem labels by @cainiao33 in https://github.com/ultralytics/ultralytics/pull/26735Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.180...v8.4.181
v8.4.181 - ultralytics 8.4.181 Fix .yml dataset discovery regression, case-insensitive source lists, non-UTF-8 hashing and auto_annotate same-stem labels (#26735)
Compare
Ultralytics 8.4.180 improves pose-aware training for YOLO26-pose and fixes several edge cases in data loading, result export, tracking, and visualizat
Ultralytics 8.4.180 improves pose-aware training for YOLO26-pose and fixes several edge cases in data loading, result export, tracking, and visualization.
.yml as well as .yaml; conversion also handles non-ASCII cache paths more reliably and skips unlabeled DOTA background images.Results.summary() now returns an empty list instead of failing when there are no boxes or oriented boxes. Validation text labels use split-relative image paths, preventing different images with the same filename from being merged into one label file.Annotator.text crash on Pillow 13 by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/26716Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.179...v8.4.180
v8.4.180 - ultralytics 8.4.180 OKS-aware label assignment for YOLO26-pose (#26695)
Compare
Ultralytics 8.4.179 focuses on export reliability and inference correctness, led by a compatibility fix for RKNN on Rockchip devices. It also improves
Ultralytics 8.4.179 focuses on export reliability and inference correctness, led by a compatibility fix for RKNN on Rockchip devices. It also improves SAM, validation, benchmarking, and training stability.
rknn-toolkit2 at versions below 2.5, because version 2.5.0 can fail to build YOLOv8 and YOLO11 models across Rockchip targets, and YOLO26 on RK3588. The RKNN smoke test now uses imgsz=64 to better catch this failure.torch.compile falls back gracefully on CUDA systems without working Triton../ entry no longer crash extraction, and transient GitLab mirror failures get one retry.Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.178...v8.4.179
Ultralytics 8.4.179 focuses on export reliability and inference correctness, led by a compatibility fix for RKNN on Rockchip devices. It also improves SAM, validation, benchmarking, and training stability.
rknn-toolkit2 at versions below 2.5, because version 2.5.0 can fail to build YOLOv8 and YOLO11 models across Rockchip targets, and YOLO26 on RK3588. The RKNN smoke test now uses imgsz=64 to better catch this failure.torch.compile falls back gracefully on CUDA systems without working Triton../ entry no longer crash extraction, and transient GitLab mirror failures get one retry.Full Changelog: v8.4.178...v8.4.179
v8.4.179 - ultralytics 8.4.179 Cap rknn-toolkit2 below 2.5 and smoke RKNN at imgsz=64 (#26727)
Compare
Release v8.4.178 makes Platform training links more reliable and brings a broad set of fixes for model validation, tracking, exports, data loading, an
Release v8.4.178 makes Platform training links more reliable and brings a broad set of fixes for model validation, tracking, exports, data loading, and training integrations. It also adds a cross-domain benchmark resource and expands Platform documentation.
Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.177...v8.4.178
Release v8.4.178 makes Platform training links more reliable and brings a broad set of fixes for model validation, tracking, exports, data loading, and training integrations. It also adds a cross-domain benchmark resource and expands Platform documentation.
Full Changelog: v8.4.177...v8.4.178
v8.4.178 - 8.4.178 Print the Platform-resolved model link after training starts (#26699)
Compare
Ultralytics 8.4.177 makes SAM prompts work consistently across image directories and video frames, alongside fixes that improve export reliability, tr
Ultralytics 8.4.177 makes SAM prompts work consistently across image directories and video frames, alongside fixes that improve export reliability, training resumes, and prediction inputs.
http are not mistaken for links.Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.176...v8.4.177
Ultralytics 8.4.177 makes SAM prompts work consistently across image directories and video frames, alongside fixes that improve export reliability, training resumes, and prediction inputs.
http are not mistaken for links.Full Changelog: v8.4.176...v8.4.177
v8.4.177 - 8.4.177 Fix SAM prompts across images and video frames (#26651)
Compare
YOLO 8.4.176 makes YOLO-World checkpoints much smaller after setting custom classes, while improving reliability across validation, fine-tuning, segme
YOLO 8.4.176 makes YOLO-World checkpoints much smaller after setting custom classes, while improving reliability across validation, fine-tuning, segmentation, tracking, and model deployment.
yolov8s-worldv2.pt, a saved checkpoint shrank from 328.5 MB to 25.8 MB, with identical predictions after reload. Caching remains available when explicitly requested.ultralytics 8.4.176 Keep CLIP out of YOLO-World checkpoints after set_classes (328 MB → 26 MB) by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/26647Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.175...v8.4.176
YOLO 8.4.176 makes YOLO-World checkpoints much smaller after setting custom classes, while improving reliability across validation, fine-tuning, segmentation, tracking, and model deployment.
yolov8s-worldv2.pt, a saved checkpoint shrank from 328.5 MB to 25.8 MB, with identical predictions after reload. Caching remains available when explicitly requested.ultralytics 8.4.176 Keep CLIP out of YOLO-World checkpoints after set_classes (328 MB → 26 MB) by @raimbekovm in #26647Full Changelog: v8.4.175...v8.4.176
v8.4.176 - ultralytics 8.4.176 Keep CLIP out of YOLO-World checkpoints after set_classes (328 MB → 26 MB) (#26647)
Compare
v8.4.175 focuses on making training, validation, inference, and deployment more reliable across existing Ultralytics models—without introducing a new
v8.4.175 focuses on making training, validation, inference, and deployment more reliable across existing Ultralytics models—without introducing a new model architecture. The headline fix prevents classification images from appearing in both training and validation splits after a split ratio changes.
uv setup and failure alerts, and the CUDA Docker image updates to PyTorch 2.14.1.dynamic=True by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/26567Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.174...v8.4.175
v8.4.175 focuses on making training, validation, inference, and deployment more reliable across existing Ultralytics models—without introducing a new model architecture. The headline fix prevents classification images from appearing in both training and validation splits after a split ratio changes.
uv setup and failure alerts, and the CUDA Docker image updates to PyTorch 2.14.1.dynamic=True by @Y-T-G in #26567Full Changelog: v8.4.174...v8.4.175
v8.4.175 - More reliable classification training and validation (#26597)
Compare
Release 8.4.174 fixes AutoBatch’s backward-pass profiling for YOLO26 and speeds up prediction summaries, alongside improvements to exports, inference,
Release 8.4.174 fixes AutoBatch’s backward-pass profiling for YOLO26 and speeds up prediction summaries, alongside improvements to exports, inference, tracking, and data handling.
save_crop file names by @MohammadHijjawi97 in https://github.com/ultralytics/ultralytics/pull/26524scale_boxes and scale_coords by @MaverickTopG in https://github.com/ultralytics/ultralytics/pull/26546ultralytics 8.4.174 Fix AutoBatch backward profiling for YOLO26 and speed up prediction logs by @aswanth-07 in https://github.com/ultralytics/ultralytics/pull/26547Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.173...v8.4.174
Release 8.4.174 fixes AutoBatch’s backward-pass profiling for YOLO26 and speeds up prediction summaries, alongside improvements to exports, inference, tracking, and data handling.
save_crop file names by @MohammadHijjawi97 in #26524scale_boxes and scale_coords by @MaverickTopG in #26546ultralytics 8.4.174 Fix AutoBatch backward profiling for YOLO26 and speed up prediction logs by @aswanth-07 in #26547Full Changelog: v8.4.173...v8.4.174
v8.4.174 - ultralytics 8.4.174 Fix AutoBatch backward profiling for YOLO26 and speed up prediction logs (#26547)
Compare
Version 8.4.173 adds native AMD Xilinx export for Versal AI Edge Series Gen 2 NPUs, opening a new path to deploy Ultralytics models on AMD edge hardwa
Version 8.4.173 adds native AMD Xilinx export for Versal AI Edge Series Gen 2 NPUs, opening a new path to deploy Ultralytics models on AMD edge hardware.
xilinx export format — Export models through AMD Quark to a quantized ONNX model and Vitis AI configuration. The format also accepts vitis, vitisai, and versal as aliases..rai model for the NPU.ultralytics 8.4.173 Add AMD Xilinx Vitis AI export for Versal AI Edge Gen 2 NPUs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/26519Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.172...v8.4.173
v8.4.173 - ultralytics 8.4.173 Add AMD Xilinx Vitis AI export for Versal AI Edge Gen 2 NPUs (#26519)
Compare
v8.4.172 makes oriented-box cropping work as expected and improves reliability across model training, export, dataset handling, and deployment.
v8.4.172 makes oriented-box cropping work as expected and improves reliability across model training, export, dataset handling, and deployment.
save_crop=True and ObjectCropper now save crops upright and aligned to each oriented bounding box. Areas extending beyond the image are filled with black.cloudpickle and filelock base dependencies and disallow new ones by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/26471Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.171...v8.4.172
v8.4.172 makes oriented-box cropping work as expected and improves reliability across model training, export, dataset handling, and deployment.
save_crop=True and ObjectCropper now save crops upright and aligned to each oriented bounding box. Areas extending beyond the image are filled with black.cloudpickle and filelock base dependencies and disallow new ones by @glenn-jocher in #26471Full Changelog: v8.4.171...v8.4.172
v8.4.172 - ultralytics 8.4.172 Add rotation-aligned OBB crops to save_crop (#26486)
Compare
v8.4.171 adds end-to-end AMD GPU support for Ultralytics workflows, alongside fixes to dataset handling, model export, and several task-specific behav
v8.4.171 adds end-to-end AMD GPU support for Ultralytics workflows, alongside fixes to dataset handling, model export, and several task-specific behaviors.
latest-amd Docker image, AMD GPU CI testing, and setup and usage guidance. MIGraphX support targets Linux x86_64 with Python 3.11 or newer.Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.170...v8.4.171
v8.4.171 - Add AMD ROCm and MIGraphX support (#24137)
Compare
One column per quarter.