ultralytics/ultralytics
Ultralytics YOLO26, YOLO11, YOLOv8 — object detection, instance segmentation, semantic segmentation, image classification, pose estimation, object tracking
How ultralytics is put together
Ultralytics is a Python computer-vision library implementing YOLO-family models (YOLOv3 through YOLO26, plus RT-DETR, SAM/FastSAM, NAS) for detection, segmentation, classification, pose estimation, tracking, and depth tasks. It is organized as a pip-installable package with an engine layer (Model/Trainer/Predictor/Validator/Exporter/Tuner), a neural-network layer (nn.tasks model construction with pluggable backends such as ONNX/OpenVINO/CoreML/Ascend/Axelera), a data layer (datasets, augmentation, loaders), trackers (BoT-SORT, ByteTrack, OC-SORT variants), a solutions layer of ready-made vision applications built on a shared BaseSolution/SolutionConfig, and a utils layer (metrics, ops, plotting, downloads). Users interact via the `yolo` CLI entrypoint (ultralytics.cfg:entrypoint) or the Python `YOLO` API; configuration flows through YAML defaults (cfg/default.yaml) merged with user overrides. Packaging is via setuptools/pyproject.toml with extensive optional dependency groups for export targets.
Languages
Frameworks
Infrastructure
Major components
ultralytics.engine
Core orchestration: Model API, trainer, predictor, validator, exporter (ONNX/TensorFlow/CoreML/OpenVINO/ExecuTorch/LiteRT etc.), and hyperparameter tuner.
ultralytics.nn
Neural network definitions and dynamic model construction (tasks.py), multi-backend inference autobackend, distillation, text models, and hardware backends (Ascend, Axelera).
ultralytics.data
Dataset loading, building, augmentation, format conversion, auto-annotation, splitting (incl. DOTA), and source loaders (images/video/streams).
ultralytics.trackers
Multi-object tracking algorithms: BoT-SORT, ByteTrack, OC-SORT, Deep OC-SORT, FastTracker, TrackTrack over a shared BaseTrack.
ultralytics.solutions
High-level vision applications (object counting, heatmaps, speed estimation, parking/queue management, blurring, cropping, workout monitoring, similarity search, Streamlit live inference, security alarm) sharing BaseSolution and SolutionConfig.
ultralytics.models
Task/model-specific packages: yolo, rtdetr, sam (incl. SAM3 build), fastsam, nas, plus LLM integration (models/llm.py) and shared loss/ops utilities.
ultralytics.cfg
Configuration system: default.yaml settings, per-dataset YAMLs, tracker config YAMLs, CLI entrypoint parsing and validation.
ultralytics.utils
Cross-cutting utilities: metrics, losses, NMS, plotting, downloads, checks, logging, torch helpers, benchmarks, Triton client, uploads, git/events helpers.
A bug-fix-heavy week with broad export improvements (LiteRT GPU delegation, ONNX opset capping, TensorRT INT8 scoping) plus many small correctness fixes across training, tracking, and data handling.
5 weeksWeek of 2026-08-24
Week of 2026-08-24latest
A bug-fix-heavy week with broad export improvements (LiteRT GPU delegation, ONNX opset capping, TensorRT INT8 scoping) plus many small correctness fixes across training, tracking, and data handling.
Week of 2026-08-17
The busiest week of the month (68 commits), headlined by correct task loading for all 20 export formats, NMS support in CoreML segment/pose exports, and a large batch of YOLOE fixes and documentation rewrites.
Week of 2026-08-10
Platform SDK exposure and LLM interface work dominated this week, alongside tracking robustness fixes and a series of small but important predictor and dataloader corrections.
Week of 2026-08-03
Depth estimation got significant attention this week (loss gating, per-image metrics, aligned postprocessing), along with YOLOE vocabulary validation, augmented-data pipeline fixes, and a security patch for shell injection.
Week of 2026-07-27
The largest week of the period (118 commits) centered on deprecating HUB in favor of Ultralytics Platform, a wave of RKNN and CoreML export fixes, and numerous visualization and tracking corrections.
Over the past month, the Ultralytics team shipped a steady stream of export and deployment fixes (TensorRT, OpenVINO, RKNN, CoreML, LiteRT) alongside a wave of bug fixes for training, tracking, and prediction edge cases. A major theme was the transition from HUB to the new Ultralytics Platform, including new SDK documentation and an LLM model interface. Performance work also landed, from faster checkpoint loading to memory reductions during ONNX INT8 export.
Week by week
2026-08-24A bug-fix-heavy week with broad export improvements (LiteRT GPU delegation, ONNX opset capping, TensorRT INT8 scoping) plus many small correctness fixes across training, tracking, and data handling.latest6 changes
Feature
GPU-accelerated end-to-end LiteRT exports
End-to-end LiteRT exports are now fully delegated to GPU accelerators for faster on-device inference.
Fix
Safer ONNX exports for CUDA
ONNX exports now cap the opset at 18 so CUDA kernels in ONNX Runtime cover all operations used.
Fix
Memory leak fix for RAM caching
Fixed a memory leak when caching images in RAM by sharing a single image buffer.
Fix
Unicode path support on Windows
Added unicode-safe image reading so non-ASCII file paths no longer break on Windows.
Refactor
Faster SAM crop encoding
SAM now encodes each image crop once and reuses it across point batches instead of re-encoding repeatedly.
Fix
Multi-dataset Ray tuning fixed
Ray hyperparameter tuning now works correctly when sweeping across multiple datasets.
2026-08-17The busiest week of the month (68 commits), headlined by correct task loading for all 20 export formats, NMS support in CoreML segment/pose exports, and a large batch of YOLOE fixes and documentation rewrites.6 changes
Fix
Correct task loading for exported models
Exported YOLO models now load with the right task type across all 20 supported export formats.
Feature
NMS built into CoreML segment and pose exports
CoreML exports for segmentation and pose models can now include non-maximum suppression directly in the exported model.
Docs
YOLOE docs rewrite
The YOLOE documentation page was rewritten and evaluation protocols were clearly labeled on LVIS result pages.
Refactor
Faster, thread-safe checkpoint loading
Loading restricted checkpoints is now about 40% faster and safe to use from multiple threads at once.
Fix
Augmentation seeds reach workers
The random seed now reaches dataloader worker processes so augmentations actually vary between training runs.
Fix
OpenVINO INT8 crash fix
Fixed a segfault when running INT8 OpenVINO inference with dynamic shapes on Intel AMX CPUs.
2026-08-10Platform SDK exposure and LLM interface work dominated this week, alongside tracking robustness fixes and a series of small but important predictor and dataloader corrections.6 changes
Feature
Standalone LLM model interface
Added a standalone LLM model interface with an OpenAI-compatible API surface, documented for users.
Feature
Platform SDK exposed from ultralytics
The Ultralytics Platform SDK is now importable directly from the main ultralytics package.
Fix
Tracking rejects invalid detections
BYTETracker now ignores zero- or negative-height detections before creating new tracks, avoiding phantom tracks.
Refactor
Faster Kalman filter math
Replaced matrix multiplications with direct slicing in the Kalman filter to speed up tracking.
Fix
Classification predict crash fixed
Fixed a crash when predicting with classification models that have no preprocessing transforms attached.
Fix
NPU turbo mode for classifiers
Enabled OpenVINO NPU_TURBO for classification models on supported Intel NPUs.
2026-08-03Depth estimation got significant attention this week (loss gating, per-image metrics, aligned postprocessing), along with YOLOE vocabulary validation, augmented-data pipeline fixes, and a security patch for shell injection.6 changes
Fix
Shell injection vulnerability patched
Fixed a security issue where untrusted requirement strings could execute shell commands via check_requirements().
Fix
Depth metrics averaged per image
Depth accuracy metrics are now averaged per image rather than pooling every pixel, giving fairer evaluation.
Fix
KITTI test split corrected
Fixed the KITTI Eigen benchmark split where three test drives had incorrectly been included in training data.
Fix
YOLOE prompt-free workflows validated
YOLOE classes are now validated up front, enabling reliable prompt-free vocabulary workflows.
Fix
OBB orientation preserved through augmentation
Rotated bounding boxes keep their orientation intact when passing through clipping augmentations.
Fix
Richer Albumentations support
Spatial Albumentations transforms are routed by type and now carry masks, polygons, and keypoints through correctly.
2026-07-27The largest week of the period (118 commits) centered on deprecating HUB in favor of Ultralytics Platform, a wave of RKNN and CoreML export fixes, and numerous visualization and tracking corrections.6 changes
Chore
HUB deprecated in favor of Platform
The old Ultralytics HUB is officially deprecated as the project consolidates around Ultralytics Platform.
Fix
RKNN export fixes
Fixed RKNN INT8 exports returning all-zero class scores and pinned package versions to make RKNN exports reliable.
Fix
TorchScript GPU device mismatch fixed
TorchScript models now run inference on the correct GPU device instead of erroring on mismatches.
Feature
YOLOE prompt embedding profiles
Added reusable YOLOE prompt embedding profiles so custom visual prompts can be saved and reused.
Refactor
Faster mask plotting
Mask drawing is now processed in row bands, cutting plotting time and memory usage significantly.
Fix
Heatmap overlay stays visible
Fixed the Heatmap visualization disappearing on video frames that contain no active tracks.
Dependencies and code review
Dependency advisories
Security Watch
No known advisories across 0 scanned dependencies.
No known advisories in the scanned dependencies.
Code review
No concrete code-level issues confirmed.
Get this report every week for your repos.
GitZoid learns each repo, reports what changed, and flags what needs attention. One flat price for the whole team.
$19 a month, flat · First 10 outputs free · No card required