NewYour coding agent can read the release notes before it upgrades.Set up the MCP server →
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11 versions withdrawn
withdrawn after publishing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.
Ultralytics v8.4.13 makes training more resilient by automatically recovering from CUDA out-of-memory (OOM) errors during the first epoch by retrying
Ultralytics v8.4.13 makes training more resilient by automatically recovering from CUDA out-of-memory (OOM) errors during the first epoch by retrying with a smaller batch size 🔁🧠🔥
_build_train_pipeline() method to rebuild loaders/optimizer/scheduler when batch size changes (used by the new OOM recovery flow).simplify=True is now forced to avoid a known runtime issue (TopK-related error in some ONNX Runtime versions).is_dgx() detection and uses it (along with Jetson JetPack 7) to trigger a TensorRT version check/reinstall path for better export reliability on those systems.setuptools<=81.0.0 to avoid breakages introduced by newer setuptools versions (notably affecting tensorflow.js export tooling).yolo26n.pt) and mention ExecuTorch/Axelera export options (documentation signposting).protobuf in the RT-DETR ONNX Runtime Python example.simplify=True for OBB export with NMS by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23580setuptools version by @Burhan-Q in https://github.com/ultralytics/ultralytics/pull/23589ultralytics 8.4.13 Retry smaller batch on training CUDA OOM by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23590Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.12...v8.4.13
Your coding agent can read these notes before it upgrades. Set up the MCP server →
Ultralytics v8.4.12 streamlines YOLOE-26 class/text-prompt handling to avoid redundant updates, while improving multispectral (grayscale) training rel
Ultralytics v8.4.12 streamlines YOLOE-26 class/text-prompt handling to avoid redundant updates, while improving multispectral (grayscale) training reliability and multi-GPU auto-selection 🧠⚡️🖥️
set_classes() when prompts already match 🧠⚡️
ultralytics/models/yolo/model.py to avoid re-running set_classes() if model.names already matches the requested classes.get_text_pe(classes)) when a real class change is needed.self.predictor.model.names = self.model.names.ultralytics/utils/autodevice.py).flags=self.cv2_flag), preventing cached .npy files from silently changing input format (ultralytics/data/base.py).model.yaml["channels"]) instead of assuming 3-channel RGB (ultralytics/utils/autobatch.py).coco8-grayscale.yaml, uses cache="disk" in grayscale tests, and cleans up cached .npy files to prevent flaky test interactions.yolo26n.onnx, yolo26n_openvino_model, etc.).ultralytics/utils/metrics.py).deque initialization in BoT-SORT tracker (ultralytics/trackers/bot_sort.py).wandb for nvidia-ml-py in the Dockerfile to better support NVIDIA GPU querying/metrics in container builds.cache="disk" and using AutoBatch on non-3-channel inputs is now more correct and consistent—reducing mismatches, mis-profiling, and possible crashes.*.npy files in grayscale tests by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23554ultralytics 8.4.12 YOLOE-26 skip set_classes if text prompts already set by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23552Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.11...v8.4.12
Ultralytics v8.4.11 makes Ultralytics Platform training uploads far more reliable by adding retryable model uploads + metrics posting, plus a new reus
Ultralytics v8.4.11 makes Ultralytics Platform training uploads far more reliable by adding retryable model uploads + metrics posting, plus a new reusable upload helper and docs 🚀📶
ultralytics/utils/uploads.py with safe_upload() that supports retries, timeouts, and optional progress bars ⏳📦_send() now retries transient failures, uses a longer timeout (30s), and does not retry most 4xx client errors (except common transient ones like timeout/rate-limit) 🛡️_upload_model() now retries signed-URL retrieval and uses safe_upload() for the actual upload to cloud storage ☁️gcsPath) as modelPath (instead of accidentally sending a local filesystem path) ✅C:\... or /home/....ultralytics 8.4.11 Platform model uploads retries by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23538Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.10...v8.4.11
Ultralytics 8.4.10 improves YOLOE-26 out-of-the-box predictions by defaulting to class-agnostic NMS, plus several export, OpenVINO, HUB, SAM, and docs
Ultralytics 8.4.10 improves YOLOE-26 out-of-the-box predictions by defaulting to class-agnostic NMS, plus several export, OpenVINO, HUB, SAM, and docs reliability upgrades 🚀🧠
agnostic_nms=True, reducing overlapping duplicate boxes across different classes (set in ultralytics/models/yolo/model.py).8.4.9 → 8.4.10.agnostic_nms (previously the argument existed but wasn’t applied in the CoreML NMS pipeline).check_tensorrt() helper and used it in export + tests to reduce missing-dependency failures; also adds special handling to pin compatible TensorRT on Jetson JetPack 7 / CUDA 13 ARM for RT-DETR exports.UnboundLocalError risk by making the inference-mode check safe when dynamic is false.HUBModelError (instead of generic ValueError or silently returning None) when loading/creating HUB models fails.names are generated (especially for single-object outputs and “visual” naming mode).segments=None during zero-area filtering, and adds a helpful Instances.__repr__() for easier debugging.model.set_classes(["person", "bus"]) (no manual text embedding handling), ExecuTorch docs updated with YOLO26 benchmarks, and Benchmark docs video link refreshed.agnostic_nms=False.agnostic_nms is enabled—fewer surprises when deploying to Apple devices.You can update with: pip install -U ultralytics ⬆️✅
test_export_engine_matrix by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/23496autobackend.py when using OpenVINO by @zboszor in https://github.com/ultralytics/ultralytics/pull/23505tensorrt on CUDA 13 ARM to 10.15.x to fix a bug with RT-DETR exports by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23499executorch doc with YOLO26 benchmarks on Raspberry Pi 5 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23500remove_zero_area_boxes instance when segment is None by @Bovey0809 in https://github.com/ultralytics/ultralytics/pull/23495ultralytics 8.4.10 YOLOE-26 default agnostic_nms=True by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23525Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.9...v8.4.10
Ultralytics v8.4.9 improves segmentation CopyPaste augmentation reliability (especially for grayscale/hyperspectral-style single-channel inputs) and s
Ultralytics v8.4.9 improves segmentation CopyPaste augmentation reliability (especially for grayscale/hyperspectral-style single-channel inputs) and strengthens the broader export/training ecosystem with better ExecuTorch + Torch support 🚀🧩
ultralytics/data/augment.py:
1 (instead of (1,1,1)), avoiding channel mismatches.8.4.8 → 8.4.9 🔢ul://... URIs (instead of silently reusing a cached .ndjson) 🧹⬇️<2.10 PyTorch upper bound in dependencies (Windows still excludes torch==2.4.0) 🔓check_executorch_requirements() to reduce platform-specific install failures 🧩compile=True, and stride calculation is simplified for more predictable behavior 🧠🔥.ndjson metadata helps ensure you train/evaluate on the latest Ultralytics-hosted dataset state when using ul://... 🔄🧪compile=True usage: better test coverage helps catch regressions for users relying on PyTorch compile for potential speedups 🧠🛠️checks.py and dist.py by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/23485compile by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23479ultralytics 8.4.9 Hyperspectral CopyPaste augmentation by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23471Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.8...v8.4.9
Ultralytics v8.4.8 makes end-to-end (NMS-free) YOLO26/YOLOv10 inference actually honor `max_det` and `agnostic_nms`, giving you predictable control ov
Ultralytics v8.4.8 makes end-to-end (NMS-free) YOLO26/YOLOv10 inference actually honor max_det and agnostic_nms, giving you predictable control over how many detections you get and how classes are handled ✅🚀
max_det + agnostic_nms (PR #23396 by @Y-T-G) 🎛️
set_head_attr(**kwargs) helper to set head/last-layer attributes like end2end, max_det, agnostic_nms 🧩end2end=True, predict and val now push max_det + agnostic_nms into the model head, so these flags actually take effect ⚙️agnostic_nms support inside the Detect head (affects top-k selection logic used in end-to-end mode) 🎯agnostic_nms into the exported head for better parity between Python and exported models 📦end2end controls across CLI/docs 📚
end2end to config/defaults and docs for predict/val/export args 🧾end2end=True/False ✅🧪end2end in metadata 🧰Path) for multi-source data configs 🗂️max_det now reliably limits outputs in end2end mode (no more “why didn’t it cap?” confusion).agnostic_nms becomes meaningful in end2end workflows, helpful when classes overlap heavily or you want “best boxes regardless of class”.If you’re using YOLO26 (the recommended model family) in end-to-end mode, this release is a practical quality-of-life upgrade. 🚀🙂
end2end mode by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23333coco12-formats predict tests by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23374select_candidates_in_gts by @PT0X0E in https://github.com/ultralytics/ultralytics/pull/23369clip package by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23368stride_val type issue in STAL by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23390rknn/executorch/paddle/imx formats by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23312torch.no_grad for torch==1.9.0 to resolve training issue by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23395close_mosaic as integer in best_hyperparameters.yaml by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23435rect=True by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23419ultralytics 8.4.8 Support max_det and agnostic_nms for end2end by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23396Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.7...v8.4.8
Ultralytics v8.4.7 adds AVIF training support and a new COCO12-Formats mini-dataset to continuously verify that *all supported image types* load corre
Ultralytics v8.4.7 adds AVIF training support and a new COCO12-Formats mini-dataset to continuously verify that all supported image types load correctly end-to-end (especially in CI) 🧪🖼️
IMG_FORMATS expanded to include avif_imread_pil) for cases where OpenCV can’t decode AVIF/HEICgenerate_coco12_formats.py) to build the set for testing/validationruns/ by default for relative projects + improves W&B traceability 🗂️runs/ by default and improve W&B run traceability by @artest08 in https://github.com/ultralytics/ultralytics/pull/23026multi_scale range sampling and clarify docs by @artest08 in https://github.com/ultralytics/ultralytics/pull/23284ultralytics 8.4.7 AVIF training and new COCO12-Formats dataset by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23358Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.6...v8.4.7
Ultralytics v8.4.6 is a reliability-focused release that fixes a multi-GPU DDP training crash and improves Ultralytics HUB/Platform dataset URI handli
Ultralytics v8.4.6 is a reliability-focused release that fixes a multi-GPU DDP training crash and improves Ultralytics HUB/Platform dataset URI handling, plus several documentation upgrades 📈🛠️
PosixPath import to the generated temporary DDP training script (ultralytics/utils/dist.py).NameError: name 'PosixPath' is not defined when model paths are passed as PosixPath objects (common on Linux) during Distributed Data Parallel training 🧩🧯ul://... dataset URLs now allow a much longer server-side preparation time (NDJSON generation) to avoid premature timeouts 🐢📦# type: comments to inline type annotations in BYTETracker (no behavior change) 🧠🔧yolo26n.pt passed as a Path/PosixPath), this release removes that blocker—less downtime, fewer confusing crashes.ul://... datasets from Ultralytics HUB/Platform 🌐✅
For more context, see the v8.4.6 release tag on GitHub.
byte_tracker.py by @ahmet-f-gumustas in https://github.com/ultralytics/ultralytics/pull/23304ultralytics 8.4.6 Add missing PosixPath import in DDP train file gen by @pfabreu in https://github.com/ultralytics/ultralytics/pull/23301Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.5...v8.4.6
Ultralytics 8.4.5 makes 2D pose `Results.summary()` safer and more compatible by correctly handling keypoints with or without visibility flags 🧍♂️✅
Ultralytics 8.4.5 makes 2D pose Results.summary() safer and more compatible by correctly handling keypoints with or without visibility flags 🧍♂️✅
Results.summary() checks kpt.has_visible and only outputs "visible" when it exists (otherwise returns just "x" and "y"), preventing crashes in mixed keypoint formats.8.4.4 → 8.4.5.hub-sdk extra from pyproject.toml (simplifies installs for users who don’t need it) 📦✂️.Results.summary() won’t fail when datasets/models omit visibility/confidence per keypoint (common in non-COCO formats).ultralytics 8.4.5 2D Pose Result.summary() support by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23293Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.4...v8.4.5
Ultralytics v8.4.4 refines MuSGD training behavior (better scaling for short vs. long runs) and includes several quality-of-life fixes for exports, se
Ultralytics v8.4.4 refines MuSGD training behavior (better scaling for short vs. long runs) and includes several quality-of-life fixes for exports, segmentation outputs, logging, and filesystem side effects 🚀🧠
🧠 MuSGD optimizer scaling update (PR #23279 by @Laughing-q)
MuSGD scale factors are now chosen more appropriately based on total training iterations:
(muon=0.1, sgd=1.0)(muon=0.5, sgd=0.5)ultralytics/engine/trainer.py.📷 Sony IMX500 export compatibility improved (PR #23266 by @Laughing-q)
🧩 Segmentation proto/output handling fixed across backends (PR #23241 by @Laughing-q)
📁 No more empty run folders when save=False (PR #23268 by @Y-T-G)
get_save_dir() no longer auto-creates directories while computing a unique run path—reduces unwanted “predict” folders appearing on disk 🧹
📈 TensorBoard OBB graph logging made safer (PR #23276 by @Y-T-G)
TensorBoard graph logging now uses smart_inference_mode() to avoid gradient tracking—often less memory/overhead and fewer callback edge cases ⚙️
🔗 Docs branding/link refresh (PR #23283 by @glenn-jocher)
Multiple repo/docs links now point to the Ultralytics Platform entry point (and wording updated accordingly) 🧭
🚀 More consistent training dynamics with MuSGD
If you use MuSGD (muon/sgd), you may see improved stability and convergence depending on whether your run is short or long—especially around the 10k-iteration threshold 🧠📉
✅ Fewer export surprises (IMX500 + segmentation exports)
IMX500 exports should fail less often due to minor layer-count differences, and segmentation outputs should be more consistent across PyTorch/exported/TF backends 📦🧩
🧹 Cleaner local runs and tooling behavior
Computing a save directory no longer creates folders prematurely, which is especially helpful for dry runs, scripts, and save=False predictions 📁✨
⚡ More robust logging during training
TensorBoard graph logging becomes less intrusive and more reliable (particularly for OBB setups) 📊🛡️
YOLOv8/YOLO11 IMX export by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23266mkdir by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23268YOLO26n for benchmark tests by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23241ultralytics 8.4.4 MuSGD optimizer scale factor update by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23279Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.3...v8.4.4
Ultralytics v8.4.3 boosts Ultralytics Platform NDJSON dataset downloads/conversion speed 🚀, improves training metric correctness 🧠, and refreshes defa
Ultralytics v8.4.3 boosts Ultralytics Platform NDJSON dataset downloads/conversion speed 🚀, improves training metric correctness 🧠, and refreshes defaults/docs around YOLO26 📚.
aiohttp only when NDJSON conversion is used (faster startup, fewer unnecessary deps) 📦8.4.2 → 8.4.3 🔖ULTRALYTICS_PLATFORM_URL to point callbacks/links to staging or local environments 🧪YOLO()/CLI fallback model becomes yolo26n.pt and many docs/examples follow suit ✅rle_loss when the model actually supports it (avoids confusing metrics) 🧾Results.summary() entries — PR #23218 by @xusuyong
ULTRALYTICS_PLATFORM_URL makes it much easier to test integrations without patching code.warmup_lr by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23234IMX inference wrapper by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23235README.md metrics table by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23238yolo26n in docs and examples by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23242ultralytics 8.4.3 Faster Platform NDJSON downloads by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23257Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.2...v8.4.3
Ultralytics v8.4.2 mainly fixes Ultralytics Platform (`ul://` / NDJSON) classification training by converting datasets into the correct on-disk layout
Ultralytics v8.4.2 mainly fixes Ultralytics Platform (ul:// / NDJSON) classification training by converting datasets into the correct on-disk layout and validating them properly, plus a few quality-of-life and docs/CI tweaks 🛠️✅
task == "classify" and creates an ImageNet-style folder layout: {split}/{class_name}/... (instead of images/ + labels/ used for detection-style tasks).data.yaml, images/{split}/, labels/{split}/).Trainer.get_dataset() now always resolves ul:// / .ndjson into a local dataset first, then runs the correct dataset validation for the task.Results.summary() now returns top-5 (PR #23215, @glenn-jocher) 🧠📋
summary() is intended to return top-5 classes + confidences instead of only top-1.save_dir reliably (PR #23191, @Y-T-G) 📁🔧
save_dir is now treated as an allowed override/config key, so setting output directories via CLI/Python overrides is more consistent.ul://...) or .ndjson exports, this release prevents failures caused by the wrong dataset folder structure and ensures the trainer checks the right dataset type.Results.summary() output in this specific tag) ⚠️save_dir without it being ignored or flagged.✅ If you’re using Ultralytics Platform + classification, v8.4.2 is a “must update”.
save_dir from argument validation by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23191ultralytics 8.4.2 Fix Platform Classify training by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23217Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.1...v8.4.2
Ultralytics v8.4.1 brings back NCNN export + inference on ARM64 (Apple Silicon/ARM servers/edge devices) 📱⚙️, plus stability fixes for pose/seg traini
Ultralytics v8.4.1 brings back NCNN export + inference on ARM64 (Apple Silicon/ARM servers/edge devices) 📱⚙️, plus stability fixes for pose/seg training 🛠️ and a big docs/benchmarks cleanup 📚✨.
ultralytics/engine/exporter.py)ultralytics/nn/autobackend.py)ultralytics/utils/benchmarks.py)proto unpacking and more robust handling when there are no foreground masks (PR #23205 and PR #23197) 🧩model.fuse() is now idempotent (won’t re-fuse an already fused model), reducing wasted time and avoiding potential side effects (PR #23189) 🔧If you want, tell me your target device (e.g., Raspberry Pi 5, Apple M2, ARM server) and whether you’re exporting YOLO11 or YOLO26, and I’ll suggest the safest export settings for NCNN ⚙️📦.
RealNVP and RLELoss by @lmycross in https://github.com/ultralytics/ultralytics/pull/23186model.fuse() statement by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23189sem_masks error and incorrect proto unwrap by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23197int8=True and update tflite wrapper by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23185PoseLoss26 to prevent crash by @Miaoge-Ge in https://github.com/ultralytics/ultralytics/pull/23205ultralytics 8.4.1 Re-enable NCNN exports for ARM64 by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23211Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.4.0...v8.4.1
Ultralytics YOLO26 has arrived. Re-engineered from the ground up by @glenn-jocher, @Laughing-q, and the Ultralytics YOLO team, YOLO26 is purpose-built
Ultralytics YOLO26 has arrived. Re-engineered from the ground up by @glenn-jocher, @Laughing-q, and the Ultralytics YOLO team, YOLO26 is purpose-built for edge and low-power environments. This release introduces a streamlined, native end-to-end NMS-free architecture, delivering faster, lighter, and more accessible deployment across all platforms.
ultralytics 8.4.0 YOLO26 Models Release by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23176Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.253...v8.4.0
v8.4.0 - New YOLO26 Models Release Latest
Latest
Compare
Ultralytics 8.3.253 adds explicit Vulkan GPU device selection for NCNN inference (plus safer PaddlePaddle/NCNN dependency handling) to improve cross-v
Ultralytics 8.3.253 adds explicit Vulkan GPU device selection for NCNN inference (plus safer PaddlePaddle/NCNN dependency handling) to improve cross-vendor GPU acceleration and reliability ⚡🖥️🛠️
device="vulkan:0" or device="vulkan:1" to choose which Vulkan-capable GPU NCNN uses (helpful on AMD/Intel/non-NVIDIA systems and multi-GPU setups).select_device() now accepts "vulkan..." device strings and leaves them unchanged so they flow cleanly into inference setup.ultralytics/nn/autobackend.py) now:
device starts with "vulkan",net.set_vulkan_device(<id>),device=torch.device("cpu") to keep the rest of the pipeline consistent (NCNN handles the GPU work internally).device=vulkan:<id>.!=3.3.0) in both export and inference paths due to an upstream breakage 🛑.Example usage (NCNN model with Vulkan):
from ultralytics import YOLO
model = YOLO("yolo11n_ncnn_model")
results = model("image.jpg", device="vulkan:0") # pick Vulkan GPU 0
ultralytics 8.3.253 Add support to select Vulkan device when using NCNN by @Faerbit in https://github.com/ultralytics/ultralytics/pull/23164Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.252...v8.3.253
v8.3.252 polishes Ultralytics console output by fixing a tqdm progress-bar issue that could print “100% complete” twice ✅📈
v8.3.252 polishes Ultralytics console output by fixing a tqdm progress-bar issue that could print “100% complete” twice ✅📈
ultralytics/utils/tqdm.py to skip the final “complete” redraw if 100% was already printed.m.shape = None) in ultralytics/engine/exporter.py to avoid stale shapes when exporting with a new image size 📦slugify() plus _get_project_name() in ultralytics/utils/callbacks/platform.py so project/run names become URL-safe (better logging links, fewer weird characters) 🌐conf=0.1 and iou=0.7 (more typical tracking behavior)conf=0.25iou=0.7True in places; FP16 half shown as False for validation docs)coco8.yaml when data isn’t providedmake_anchors() in ultralytics/utils/tal.py iterates over len(feats) instead of directly iterating a strides tensor to avoid TracerWarning during tracing 🛠️imgsz, improving export consistency across runs 🔁ultralytics 8.3.252 Fix TQDM duplicate 100% states by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23158Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.251...v8.3.252
v8.3.251 improves training/integration reliability by initializing Trainer callbacks earlier (so Ultralytics HUB/Platform sees the original data input
v8.3.251 improves training/integration reliability by initializing Trainer callbacks earlier (so Ultralytics HUB/Platform sees the original data input like ul://...) while also polishing profiling accuracy, tuning stability, and device/dataset/docs support 🧩🚀
on_pretrain_routine_start now runs before dataset resolution (get_dataset()), keeping the original args.data intact (e.g., ul:// URIs).modelId for better event correlation.ul:// weights loading 🛰️
YOLO(...)._load() now recognizes ul:// as a valid remote source prefix (alongside http(s)://, rtsp://, etc.).model.info() now accepts an imgsz argument, and benchmarks use it so FLOPs reflect the actual input size (not always 640).Tuner now passes save_dir into subprocess runs to ensure consistent output paths.run_ray_tune) 🧹is_rockchip() now handles SoC strings with suffixes like rk3588-..., improving detection on more boards.ul://user/datasets/name) before it’s resolved/rewritten—improving traceability and “training_started” metadata accuracy.imgsz.run_ray_tune by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23147is_rockchip function to handle more device variants by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23153ultralytics 8.3.251 Earlier trainer callback init by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23155Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.250...v8.3.251
Ultralytics v8.3.250 adds out-of-the-box support for the TT100K traffic sign dataset 🚦—plus smoother Ultralytics HUB dataset handling, better run dire
Ultralytics v8.3.250 adds out-of-the-box support for the TT100K traffic sign dataset 🚦—plus smoother Ultralytics HUB dataset handling, better run directory behavior, and a few quality/build/docs fixes 🛠️📚
ultralytics/cfg/datasets/TT100K.yaml with 221 classes and full dataset metadata.ul://.../datasets/... sources by resolving them locally first, then performing NDJSON → YOLO conversion automatically.--project path handling for “nested” names like user/project, placing outputs under the normal runs directory.ConfusionMatrix(names=...) default from [] to {} to match its annotated type and avoid mutable-default pitfalls.ONNXRUNTIME_ROOT in CMake (so custom install paths work as documented).urllib3 in the RT-DETR ONNXRuntime Python example requirements.yolo detect train data=TT100K.yaml model=yolo11n.pt 🏁ul://), training should now “just work” even when the dataset starts as NDJSON—less manual conversion and fewer path surprises 📦✅--project values won’t accidentally create confusing folder structures; runs stay grouped under the expected task directory 🗂️ConfusionMatrix default parameter by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23138ultralytics 8.3.250 Tsinghua-Tencent 100K dataset by @PrashantDixit0 in https://github.com/ultralytics/ultralytics/pull/22892Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.249...v8.3.250
Ultralytics 8.3.249 adds an official NVIDIA ARM64 Docker image for running YOLO11 on JetPack 7 / ARM64 GPUs (plus a few usability + contributor/docs i
Ultralytics 8.3.249 adds an official NVIDIA ARM64 Docker image for running YOLO11 on JetPack 7 / ARM64 GPUs (plus a few usability + contributor/docs improvements) 🚀🐳
docker/Dockerfile-nvidia-arm64 based on NVIDIA’s PyTorch NGC image (nvcr.io/nvidia/pytorch:25.10-py3)ultralytics/ultralytics:latest-nvidia-arm64opencv-python-headless (avoids GUI dependency issues in containers) 🧩onnxruntime_gpu + reinstalls torch/torchvision for CUDA compatibility 🔥yolo11n.pt) 📦ul:// URIs (Ultralytics Platform / HUB-style referencing)
resolve_platform_uri() resolves ul://... into signed download URLs (using ULTRALYTICS_API_KEY)check_file() now accepts ul://... and automatically downloads/caches the file locally 📥🗂️ObjectCounter.display_counts so show_in / show_out flags behave correctly (prevents empty/incorrect labels) 🛠️CONTRIBUTING.md with PR size/scope guidance, feature PR expectations, and automated review notes to streamline contributions 🤝ul://
If you want to try the new ARM64 image on JetPack 7:
t=ultralytics/ultralytics:latest-nvidia-arm64
sudo docker pull $t && sudo docker run -it --ipc=host --runtime=nvidia $t
CONTRIBUTING.md by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23133ObjectCounter.display_counts by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23132ultralytics 8.3.249 Jetson AGX Thor and DGX Spark Docker by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23111Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.248...v8.3.249
Ultralytics v8.3.248 makes auto-installation more reliable by ensuring uv pip install installs into the *currently running Python environment* (plus a
Ultralytics v8.3.248 makes auto-installation more reliable by ensuring uv pip install installs into the currently running Python environment (plus a couple of small quality-of-life fixes) 🧰✅
check_requirements() runs uv pip install with --python {sys.executable} so installs go to the right environment (venv/conda/system).uv pip install --system, which could accidentally install into the system Python.imxconv-pt discovery 🔍📦
imxconv-pt in the current Python’s bin/ directory first (venv-friendly), then falls back to PATH.pip install imx500-converter[pt].TORCH_CPP_LOG_LEVEL=ERROR to suppress noisy PyTorch NNPACK warnings in containers.VIRTUAL_ENV isn’t set) should see more predictable installs and fewer broken environments.imxconv-pt) should “just work” more often inside venvs, and fails faster with an actionable message when the tool isn’t installed.plot_images by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23114ultralytics 8.3.248 Target correct Python env for auto-install by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23118Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.247...v8.3.248
Ray Tune + Weights & Biases (W&B) tuning runs now get unique, trial-specific names in Ultralytics 8.3.247, making hyperparameter sweeps much easier to
Ray Tune + Weights & Biases (W&B) tuning runs now get unique, trial-specific names in Ultralytics 8.3.247, making hyperparameter sweeps much easier to track 🧪📊
train 🏷️
trial_id to append a unique suffix{base_name}_{trial_suffix} (example: my_run_00000) ✅if not preds: return)Boxes.xywh) and clarifies tensor dimension naming in transformer codeultralytics 8.3.247 Improve Ray Tune trial names logged to W&B by @raimbekovm in https://github.com/ultralytics/ultralytics/pull/23084Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.246...v8.3.247
Ultralytics v8.3.246 upgrades training reporting for Ultralytics HUB by uploading rich, interactive plot data + class names at the end of training, ma
Ultralytics v8.3.246 upgrades training reporting for Ultralytics HUB by uploading rich, interactive plot data + class names at the end of training, making results easier to explore and understand 📈🏷️✨
on_train_end(), Ultralytics now collects plots from both the trainer and validator and sends them with the final "training_complete" event (e.g., confusion matrix, PR curves, metric–confidence curves)."training_complete" message now includes classNames, so plots/results can be labeled correctly on the frontend.on_plot) now pass raw arrays + plot type, enabling interactive rendering in Ultralytics HUB:
check_requirements() now surfaces more detail (e.g., captured command output) so dependency issues are easier to debug.onnxslim minimum bumped to >=0.1.82 to reduce ONNX export tooling issues.ruamel.yaml<0.19.0 to prevent version-related export failures.For the full context, see PR #23104 by @glenn-jocher and related PRs included in the v8.3.246 release notes.
check_requirements() by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23097ultralytics 8.3.246 Save training plots data for loggers by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23104Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.245...v8.3.246
Ultralytics 8.3.245 enhances training/run telemetry by reporting system hardware capacity (RAM + disk) and best model file size, plus a small docs foo
Ultralytics 8.3.245 enhances training/run telemetry by reporting system hardware capacity (RAM + disk) and best model file size, plus a small docs footer refresh 🖥️📈📦
_get_environment_info() now reports:
psutilshutil.disk_usage("/")on_train_end(), the callback records the best checkpoint file size (modelSize) and includes it in the final training-complete payload..view(-1) / .reshape(...) in attention/query-selection code comments (no runtime behavior change).ultralytics 8.3.245 2026 Docs footer and System metrics updates by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23095Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.244...v8.3.245
YOLO inference is now safer when switching devices mid-run by automatically resetting the predictor to avoid stale state issues 🔄🖥️✅
YOLO inference is now safer when switching devices mid-run by automatically resetting the predictor to avoid stale state issues 🔄🖥️✅
ultralytics/engine/model.py, the model now detects when device= changes between predict() calls (e.g., CPU ↔ GPU) and re-initializes self.predictor instead of reusing the old one.setuptools<71.0.0 workaround and replaces it with a targeted fix: on Linux + ARM64 + Docker, require packaging>=22.0 to avoid build/export issues.model.predict(device="cpu") and other times device=0, the predictor will no longer “carry over” device-specific state that can cause incorrect or surprising results.ultralytics 8.3.244 Reset YOLO predictor on device change by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/23086Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.243...v8.3.244
v8.3.243 sharpens Ultralytics training telemetry and console output—cleaner progress bars, richer run metadata for Ultralytics Platform/HUB-style trac
v8.3.243 sharpens Ultralytics training telemetry and console output—cleaner progress bars, richer run metadata for Ultralytics Platform/HUB-style tracking, and safer/less noisy logging 🚀🧹📡
_get_environment_info() collector (OS, Python, hostname, CPU/GPU details, command, and Git repo/branch/commit when available) plus model info (params, GFLOPs, class count) into the training-start event payload.pip install ultralytics-opencv-headless for servers/CI/Docker to avoid OpenCV GUI libGL issues.docker run commands now include --runtime=nvidia to reduce “GPU not found” surprises.ULTRALYTICS_SKIP_REQUIREMENTS_CHECKS=1 skips requirement checks/auto-install behavior (useful for controlled environments)."" as a background class may help in some cases.--runtime=nvidia usage description to Dockerfiles by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23080ultralytics 8.3.243 Deduplicate ConsoleLogger progress bars by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23082Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.242...v8.3.243
Ultralytics 8.3.242 adds a new headless OpenCV distribution (ultralytics-opencv-headless) to make installs on servers/CI/containers smoother and more
Ultralytics 8.3.242 adds a new headless OpenCV distribution (ultralytics-opencv-headless) to make installs on servers/CI/containers smoother and more reliable 📦🖥️
ultralytics-opencv-headless (PR #23075 by @glenn-jocher)
ultralytics, swapping opencv-python → opencv-python-headless during the build.1.23.5 to avoid TensorRT export breakage (PR #23032) 🔧masks2segments() now accepts NumPy arrays or PyTorch tensors, with safer dtype/contiguity handling for OpenCV contours (PR #23025) ✅fraction to Axelera export args and ensures a default calibration dataset when missing (PR #23045) ⚙️v8OBBLoss filtering of tiny rotated boxes (PR #23061) 🎯dfl), plus small code cleanups/staticmethod refactors (multiple PRs) 📚✨ultralytics-opencv-headless avoids GUI-related OpenCV dependencies, reducing common install/runtime issues on Linux servers, Docker, and CI.masks2segments() works cleanly with both NumPy and Torch pipelines, with fewer OpenCV edge-case failures.If you want, I can also suggest when to install ultralytics vs ultralytics-opencv-headless based on your environment (desktop vs Docker/CI).
masks2segments to support numpy input by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23025numpy for JetPack 5 Docker builds by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23032numpy pin for Sony IMX export and JetPack 6 systems by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/23030sam-3.md docs by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/23041box_loss to cls_loss in TVPDetectLoss by @ShuaiLYU in https://github.com/ultralytics/ultralytics/pull/23046ultralytics 8.3.242 New ultralytics-opencv-headless PyPI package by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/23075Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.241...v8.3.242
v8.3.241 boosts ONNX Runtime performance on macOS by preferring CoreML on Apple hardware, stabilizes ExecuTorch exports, and expands hardware/image su
v8.3.241 boosts ONNX Runtime performance on macOS by preferring CoreML on Apple hardware, stabilizes ExecuTorch exports, and expands hardware/image support for Rockchip and 2‑channel data — plus a round of documentation and tooling polish. 🍏🚀
Mac ONNX Runtime: Prefer CoreML on Apple hardware (@glenn-jocher)
device="mps", Ultralytics now prefers CoreMLExecutionProvider over CPU for ONNX Runtime.ExecuTorch export reliability improvements (@glenn-jocher, @onuralpszr)
export_executorch() now pins numpy<=2.3.5 to avoid known coremltools failures with newer NumPy.import torch in the ExecuTorch export path for a cleaner exporter.New hardware support: Rockchip RV1126B (@venjye, @glenn-jocher)
"rv1126b" to supported Rockchip chips in code.Better image handling: 2‑channel images in Annotator (@kenanking)
ultralytics.utils.plotting.Annotator now correctly handles 2‑channel numpy images by converting them to 3‑channel before drawing and saving.TF/SavedModel & TensorRT backend cleanups (@glenn-jocher)
tf.saved_model.load for this path.SAM 3 docs: correct save usage (@dingjie-ai)
SAM3SemanticPredictor examples now set save=True through the overrides dict at initialization, not per call.Sony IMX500 & solutions docs polish (@glenn-jocher)
YOLOSegment example for the Sony IMX500 integration docs.Docs site & authorship updates (@glenn-jocher, @RizwanMunawar, @onuralpszr)
ultralytics/llm@v0.1.8 for a better in-docs chat experience.mkdocs-ultralytics-plugin dev dependency to >=0.2.4.Faster inference on macOS with ONNX Runtime 🍏⚡
device="mps" now automatically benefit from Apple Neural Engine / GPU acceleration via CoreML when available.More reliable ExecuTorch exports 📦🧩
Broader deployment options on edge devices 🧱🚀
Smoother visualization for non-standard data 🌈🖼️
Cleaner and more predictable backends 🧠🔧
Clearer documentation & better tooling experience 📚✨
save=True should be configured.Overall, v8.3.241 is a backend and compatibility–focused release: faster ONNX Runtime on macOS, more dependable exports, better edge-device coverage, and smoother visualization for specialized image formats — with documentation and tooling kept in sync. 🎛️🚀
ultralytics 8.3.241 ORT CoreML execution provider when device="mps" by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22984Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.240...v8.3.241
ultralytics 8.3.240 is a stability-focused release that makes SAM 2/3 video segmentation more reliable (especially on Apple Silicon), tightens GPU mem
ultralytics 8.3.240 is a stability-focused release that makes SAM 2/3 video segmentation more reliable (especially on Apple Silicon), tightens GPU memory usage, and polishes exports and documentation for smoother day‑to‑day workflows. 🚀
🔧 SAM3 on Apple Silicon (MPS) now runs reliably
repeat() on complex tensors.torch.view_as_real() / torch.view_as_complex() for an MPS-safe path while keeping existing behavior on CUDA/CPU.🧠 SAM2/SAM3 video memory usage is now bounded
_prune_non_cond_memory() in SAM2 video predictors to periodically clear “non-conditioning” frame outputs.num_maskmem, memory_temporal_stride_for_eval) and runs automatically during video inference.🎥 SAM3 semantic video tracking is faster and cleaner
init_trk_keep_alive and max_trk_keep_alive to 30) while decreasing keep-alive when masks are empty.📦 ONNX export environments are more complete
onnxslim>=0.1.80 to the export extra in pyproject.toml, so pip install ultralytics[export] pulls in onnxslim automatically.📥 Clearer setup for SAM 3 model weights
sam3.pt download link.💬 Docs chat widget is now easier to extend
UltralyticsChat instance in a variable (const ultralyticsChat = ...) so future scripts can programmatically control the widget.🖼️ Cleaner docs for color previews & Markdown-heavy pages
docs/build_reference.py now preserves Markdown structures (tables, admonitions, code blocks, headings, lists) instead of collapsing lines.Colors class docstring so the Markdown table reads cleanly, with “Attributes” and “Examples” sections following it.🔁 Branding & integration docs refresh for Comet
🍎 Better Apple Silicon experience
🧹 More stable long video runs
🎯 More predictable tracking behavior
🛠️ Smoother export and deployment workflows
onnxslim in the export extras reduces “missing dependency” surprises for users who export to ONNX and then run post‑processing or optimization.📚 Easier SAM 3 onboarding
sam3.pt from Hugging Face mean fewer setup issues and less guesswork for new SAM 3 users.📘 More maintainable and readable documentation
utils.plotting.Colors() previews by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22973SAM3SemanticVideoPredictor speed issue by removing unmatched objects by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/22976ultralytics 8.3.240 SAM3 Apple MPS tensor repeat() fixes by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22968Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.239...v8.3.240
Ultralytics v8.3.239 focuses on smoother SAM 3 text prompting, richer system monitoring via SystemLogger(), quieter logs, and broader ONNX export supp
Ultralytics v8.3.239 focuses on smoother SAM 3 text prompting, richer system monitoring via SystemLogger(), quieter logs, and broader ONNX export support (including Jetson JetPack 6), plus improved contributor attribution in the docs. 🚀
SAM 3 text prompting made plug‑and‑play 🧩
SimpleTokenizer().bpe_simple_vocab_16e6.txt.gz.SAM3SemanticPredictor and SAM3VideoSemanticPredictor no longer take a bpe_path argument.clip isn’t installed, Ultralytics auto-installs it from the Ultralytics CLIP repo.SystemLogger() gains real‑time throughput metrics 📈
SystemLogger.get_metrics() now supports rates=True to report disk and network MB/s:
read_mbs, write_mbsrecv_mbs, sent_mbsrates=False) still returns cumulative MB since logger init."0", "1").Centralized warning suppression for cleaner logs 🔕
warnings.filterwarnings(...) calls into ultralytics/utils/__init__.py.ultralytics/engine/exporter.py)ultralytics/nn/text_model.py)ultralytics/utils/plotting.py)ultralytics/utils/callbacks/tensorboard.py)ONNX export updated to support ONNX 1.20.0+ 📦
onnx>=1.12.0,<2.0.0 instead of capping at 1.19.1 for:
export_onnxexport_saved_model (TensorFlow route)pyproject.toml updated to note MacOS testing up to onnx==1.20.0.Docs & authorship improvements 📚
SAM3SemanticPredictor / SAM3VideoSemanticPredictor without bpe_path.mkdocs.yml nav updated to drop the removed tokenizer_ve reference page.docs/mkdocs_github_authors.yaml expanded with many new GitHub usernames and avatars, plus aliases and Ultralytics team mappings for better contributor attribution.CI / tooling maintenance 🧰
actions/download-artifact bumped from v6 → v7 (Node.js 24 runtime) in publish.yml.8.3.239 in ultralytics/__init__.py for clear release tracking.Easier SAM 3 text-based segmentation 💬
bpe_path wiring.sam3.pt and go.Better monitoring for training and deployment 📊
SystemLogger.get_metrics(rates=True) lets you watch live I/O throughput, which is especially helpful when:
Quieter, more focused logs 🤫
More robust ONNX export, including Jetson 🖥️
<2.0.0 (including 1.20.0) unblocks environments like Jetson JetPack 6 that depend on newer ONNX.Improved documentation and community recognition 🤝
Overall, v8.3.239 is a quality‑of‑life release: SAM 3 is simpler to use, monitoring is more informative, logs are tidier, and ONNX export is more future‑proof—without breaking existing workflows. ✅
CLIP tokenizer functionality and remove the need of bpe_simple_vocab for SAM3 by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22958SystemLogger() improvements by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22965Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.238...v8.3.239
Large pass standardizing terminology (pretrained vs pre-trained), cleaning grammar, clarifying Explorer’s deprecated status and pointing users to Ultr…
Ultralytics v8.3.238 is a refinement-heavy release that makes SAM3 concept/video segmentation more robust, faster to work with, and easier to install, while also stabilizing model export (especially TFLite/ONNX) and polishing docs & CI workflows. 🧠🧩
SAM3 architecture refactors (core of this release) 🧱
sam_forward_feature_levels) reduces duplicated code and unifies how multi-scale features + positional encodings are produced._prepare_backbone_features), centralizing multi-prompt batching and reducing redundant computation.SAM3 stability & correctness fixes 🔧
SAM3SemanticPredictor no longer corrupts its cached backbone features when you query the same image with different numbers of text prompts—fixes shape mismatch crashes in “encode once, query many times” workflows.ftfy, regex, iopath) are auto-checked and prompted for install instead of crashing with ModuleNotFoundError.from ultralytics.models.sam import ...).SAM3 video API & docs improvements 🎥
from ultralytics.models.sam import SAM3VideoSemanticPredictor.v8.3.237.Export & deployment reliability 🛫
--half and --format=onnx and device=cpu are all true—prevents ONNX-specific code running for other formats and causing avoidable errors.half=True with nms=True.torch<1.10, and the exporter now asserts torch>=1.10 for MNN to avoid runtime segmentation faults.onnx>=1.12.0,<=1.19.1 with comments that this remains until onnx_graphsurgeon supports newer ONNX.CI, tooling & infra 🏗️
SlowTests, and Slack alerts will also trigger if SlowTests fail—making flaky or long-running issues more visible.apt-get update, installs libicu-dev, and clears apt lists to reduce image bloat.actions/upload-artifact bumped from v5 → v6 (Node 24 runtime).Examples & ecosystem tweaks 🌍
.gitignore now allows examples/**/requirements.txt; new example-specific requirements were added (e.g., for ONNXRuntime/OpenCV demos).save_video=False by default).Docs & UX polish 📚
pretrained vs pre-trained), cleaning grammar, clarifying Explorer’s deprecated status and pointing users to Ultralytics HUB.More reliable SAM3 workflows for research & production 🧠🚀
Smoother out-of-the-box experience for SAM3 users 💡
from ultralytics.models.sam import ...) make it easier to copy examples and reduce breakage if internals move.Safer, more predictable model export across formats 📦
Higher CI signal quality & maintainability ⚙️
Clearer documentation for a wide audience 📖✨
SAM3VideoSemanticPredictor by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/22945half=True and nms=True by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22946ultralytics 8.3.238 Refactor SAM3 forward convolutions by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22942Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.237...v8.3.238
Ultralytics 8.3.237 adds full SAM 3 image & video segmentation support (including text & exemplar prompts and tracking), improves export behavior (ONN
Ultralytics 8.3.237 adds full SAM 3 image & video segmentation support (including text & exemplar prompts and tracking), improves export behavior (ONNX FP16 on CPU, Edge TPU/IMX deps), and polishes training, validation, and docs for smoother day‑to‑day use. 🚀
🧠 SAM 3 integration (image & video)
build_sam3.py) with ViT backbone, transformer encoder/decoder, text encoder, geometry encoders, and video tracker (SAM3Model, SAM3SemanticModel and SAM3-specific modules).sam3.pt and builds the SAM 3 tracker via build_interactive_sam3.🎛️ New SAM 3 predictors & APIs
SAM3Predictor – SAM3-style interactive segmentation.SAM3SemanticPredictor – text & exemplar based concept segmentation on images.SAM3VideoPredictor – video tracking with box prompts.SAM3VideoSemanticPredictor – video concept tracking (text + boxes + masklets).ultralytics.models.sam.__all__ and SAM’s task_map, so SAM("sam3.pt") routes to the right predictor.🧩 SAM pipeline upgrades (SAM / SAM2 / SAM3)
Predictor.setup_source now accepts an explicit stride, and SAM/SAM2/SAM3 predictors use it to enforce square image sizes and consistent feature shapes.MemoryEncoder and MaskDownSampler (interpolation to fixed sizes, higher-res mask handling).get_abs_pos, concat_rel_pos).SAM2Model.set_imgsz generalized (no longer hardcoded stride 16) and specialized SAM3Model added with improved mask post-processing and non‑overlap suppression.🖼️ SAM 3 docs & usage examples
docs/en/reference/models/sam/sam3/* for all SAM3 modules (encoder, decoder, geometry encoders, text encoder, tokenizer, etc.).docs/en/models/sam-3.md rewritten from “API preview” into concrete usage:
sam3.pt from the SAM 3 repo.bpe_simple_vocab_16e6.txt.gz) for text prompts.SAM3SemanticPredictor)SAM3VideoPredictor)SAM3VideoSemanticPredictor)SAM("sam3.pt") while clarifying the difference vs. concept segmentation.🔢 ONNX FP16 export on CPU
half=True on CPU applies only to GPU TorchScript export.half=True on CPU:
onnxruntime.transformers.float16.convert_float_to_float16(keep_io_types=True).🐧 Edge TPU & IMX export dependency management
export_edgetpu: shell apt-get calls replaced with centralized check_apt_requirements(["edgetpu-compiler"]).export_imx: Java installs now use check_apt_requirements() for:
openjdk-21-jre on Ubuntu / Debian Trixie.openjdk-17-jre on Raspberry Pi / Debian Bookworm.check_apt_requirements() now runs apt update with check=True, surfacing update failures instead of silently ignoring them.🔄 More flexible resume‑training overrides
save_period, workers, cache, patience, time, freeze, val, plots.📏 RT-DETR validation scaling fix
ratio_pad injection and replaced with a clean Compose([]).scale_preds() override to make scaling behavior explicit and safe for future changes.🧭 OBB plotting robustness
OBBValidator.plot_predictions() reworked to:
plot_images() directly, avoiding redundant xywh2xyxy conversions and mismatched formats.📚 Data augmentation docs: scale range clarified
scale hyperparameter doc updated from “≥ 0.0” to 0.0 - 1.0 in the guide and macro tables, matching real‑world usage and preventing unstable configs.🧠 Richer segmentation capabilities with SAM 3
🧪 More predictable, robust SAM/SAM2/SAM3 behavior
stride handling avoids subtle spatial shape bugs and mismatches in encoders/decoders.🚀 Better export experience across devices
🧪 Easier experiment management & training control
🎯 More reliable evaluation & visualization
ratio_pad hacks and is prepared for future scaling logic via scale_preds().📖 Clearer documentation & safer configs
scale augmentation bounds (0–1) help avoid extreme settings that could degrade training stability or accuracy.check_apt_requirements function by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/22925predictions.json for RTDETR by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22817ultralytics 8.3.237 SAM3 integration by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/22897Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.236...v8.3.237
Ultralytics 8.3.236 adds first-class Axelera Metis AIPU export and inference support for YOLO models, plus several documentation, integration, and too
Ultralytics 8.3.236 adds first-class Axelera Metis AIPU export and inference support for YOLO models, plus several documentation, integration, and tooling improvements that make deployment, tracking, and docs navigation smoother. 🚀
🔥 New Axelera export format for Metis AIPU
axelera export target in the exporter:
yolo export model=yolo11n.pt format=axeleramodel.export(format="axelera")yolo11n_axelera_model/ containing:
.axm modelmetadata.yaml with classes, image size, etc.🧠 Axelera compilation & quantization pipeline
.axm.CompilerConfig presets for YOLO11 vs YOLOv8 for stability and performance.int8=True for mixed-precision AIPU export.opset=17 for ONNX export to satisfy compiler requirements.data=coco128.yaml for Axelera if not provided.⚙️ Axelera runtime backend in AutoBackend
*_axelera_model/ path..axm file inside the directory.axelera.runtime.op.load(...).metadata.yaml to keep YOLO’s normal postprocessing pipeline.yolo predict model=yolo11n_axelera_model source=...YOLO("yolo11n_axelera_model")(source)🧪 CI & tests for Axelera support
test_export_axelera:
🧩 New system-level & Python helpers
check_apt_requirements():
apt packages (via dpkg -l) and attempts sudo apt install -y ....IS_PYTHON_3_10 to gate Axelera export.TORCH_2_8 for version checks.📚 Axelera integration docs upgraded from “coming soon” to usable
docs/en/integrations/axelera.md completely rewritten:
apt repo setup.model.export(format="axelera") / yolo export ...yolo11n_axelera_model.🧠 New Neptune integration documentation
docs/en/integrations/neptune.md and link wired into Integrations index.project= slug usage (e.g. workspace/name).📺 Updated prediction tutorial video
predict mode docs now embed a new YouTube tutorial focused on:
🧭 Docs navigation, search, and localization improvements
search plugin set to enabled: false.build_docs.py now deletes search.json from the built site./sitemap.xml requests (from Weglot hreflang handling) and returns an empty sitemap to avoid 404s.hreflang attributes to detect languages.chat.min.js bumped to v0.1.6 via CDN for a smoother embedded Ultralytics LLM experience.🧪 Example scripts hardening
ort.get_available_providers() and choose CUDAExecutionProvider and/or CPUExecutionProvider as available.np.array(...).flatten() and explicit int casting to avoid index shape issues.--img argument instead of a hardcoded image path.urllib3 in RTDETR ONNX example bumped to 2.6.0 for up-to-date HTTP stack.Unlock high-performance edge deployment on Axelera Metis ⚡
export(format="axelera").Reduce friction in environment setup 🛠️
check_apt_requirements() and structured version checks help users automatically install system dependencies and validate Python/Torch versions needed for Axelera.Keep documentation aligned with real-world usage 📘
Improve docs browsing and internationalization UX 🌍
Make examples more robust across environments 💻
Overall, 8.3.236 is a deployment‑focused release: it brings YOLO into the Axelera Metis ecosystem, strengthens experiment tracking and documentation UX, and polishes supporting tooling so users can deploy and iterate more confidently. ✨
ultralytics 8.3.236 add Axelera export for YOLO on Metis AIPU by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/22802Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.235...v8.3.236
Ultralytics v8.3.235 makes IMX500 exports more powerful (now with YOLO11 instance segmentation), modernizes export backends (CoreML, ExecuTorch, ONNX/
Ultralytics v8.3.235 makes IMX500 exports more powerful (now with YOLO11 instance segmentation), modernizes export backends (CoreML, ExecuTorch, ONNX/Torch), and improves logging and Jetson stability—while lifting an older PyTorch pin so you can safely use newer versions. 🚀
🧠 Sony IMX500: Full YOLO11 instance segmentation support
segment_forward path and segmentation handling in ultralytics.utils.export.imx.Exporter.__call__ to accept segmentation models and enforce int8=True and nms=True for segment too.📝 IMX500 docs & deployment examples upgraded
yolo11n-seg_imx_model.🔄 IMX export robustness & dependency updates
sony-custom-layers to edge-mdt-cl for custom layers/NMS.imx500-converter[pt]>=3.17.3, edge-mdt-cl<1.1.0, edge-mdt-tpc>=1.2.0, model-compression-toolkit>=2.4.1, pydantic<=2.11.7.onnx_export_patch() context to work around PyTorch 2.9+ ONNX export issues.edge-mdt-cl NMS ops and onnxruntime-extensions.🔁 CoreML & ExecuTorch export/inference modernized
coremltools version raised from >=8.0 to >=9.0.numpy>=1.14.5,<=2.3.5 for CoreML to avoid breakage with newer numpy prereleases.pyproject.toml export extrasexport_coreml in the exporter1.0.0 to 1.0.1 for both export and inference paths.setuptools<71.0.0 to avoid known compatibility issues.🧪 IMX export test re-enabled and tightened
test_export_imx now:
MODEL constant instead of hard‑coded yolov8n.pt.🐍 Python export Docker image: PyTorch pin removed (current PR focus)
torch<=2.8.0 in pyproject.toml.numpy==1.26.4 for Sony IMX export stability.onnx_export_patch() helper ensures ONNX export behaves correctly with Torch 2.9+ by temporarily disabling Dynamo during export.🚀 Jetson JetPack 6 Docker: safer ONNX version
Dockerfile-jetson-jetpack6 now patches pyproject.toml to enforce onnx>=1.12.0,<1.20.0.onnx 1.20.0 on Jetson JetPack 6.📈 ClearML & Neptune logging made future‑proof
trainer.plots and trainer.validator.plots."batch" (debug/per‑batch images).results.png, confusion_matrix.png, etc.—so new/custom plots are automatically logged.🔧 Dependency installation behavior hardened
attempt_install no longer passes --prerelease=allow to uv pip install.📚 New reference docs entries
ultralytics.utils.export.imx.segment_forwardultralytics.utils.patches.onnx_export_patch🔖 Version bump
__version__ updated from 8.3.234 to 8.3.235.✅ Better Sony IMX500 support out of the box
⚙️ Modern, stable export backends
🧪 More reliable ONNX & Torch 2.9+ workflows
<=2.8.0 pin in the export Docker while adding onnx_export_patch() means:
🛡️ Safer Jetson deployments
onnx <1.20.0 for JetPack 6 avoids known TFLite export bugs.📊 Richer experiment tracking with minimal maintenance
🧱 More predictable environments
uv decreases surprise breakages.imx500-converter 3.17.3 release and Segmentation support by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/22146Dockerfile-jetson-jetpack6 to pin onnx<1.20.0 by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/22864ultralytics 8.3.235 Remove torch 2.8.0 pin for Sony IMX export by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/22874Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.234...v8.3.235
ultralytics 8.3.234 Security fix: ast.literal_eval for safer evaluation of metadata strings by @onuralpszr in https://github.com/ultralytics/ultralyti…
Ultralytics 8.3.234 is a small but important maintenance release focused on safer model metadata handling, more robust training augmentations, and smoother docs & tooling—with no breaking changes for typical YOLO users. 🚀
🔐 Safer metadata parsing in model export (current PR #22847)
eval with ast.literal_eval when reading string metadata (e.g. imgsz, names, kpt_shape, kpt_names, args) in torch_to_mnn inside the autobackend.eval calls in cfg2task with # nosec B307 to clearly document that they are safe, controlled uses on known attributes.🧪 More robust Albumentations label handling (#22846)
labels["cls"] from Albumentations transforms is always stored as a 2D column array (num_boxes, 1) instead of a 1D array.🐍 Dependency & Python version alignment for dev installs (#22830)
zensical dev dependency to require Python 3.10+ (matching what the docs tooling already expects).🎥 Improved segmentation docs with a video tutorial (#22825)
🤖 Upgraded in-docs LLM chat widget (#22832, #22845, #22861)
mkdocs.yml from v0.0.8 → v0.0.9 → v0.1.0 → v0.1.2.✨ JavaScript & CSS cleanups in docs UI (#22842, #22844)
Number.parseInt / Number.parseFloat, Number.isNaN, and cleaner regexes.!important flags and clarifying language switcher and banner styles.🔗 Documentation & link updates (#22823, #22829)
onnxruntime-gpu wheel install to a clean Ultralytics GitHub asset link, making commands easier to copy and more reliable.🔢 Version bump
8.3.233 → 8.3.234.🛡️ Higher security when loading/exporting models
eval with ast.literal_eval significantly reduces the risk of arbitrary code execution from crafted metadata strings (e.g. in exported models).⚙️ More stable training with Albumentations
(num_boxes, 1) shape for class labels avoids hidden bugs and crashes in data pipelines.👨💻 Smoother development experience
zensical in dev extras means fewer dependency resolution headaches for contributors and power users working on the repo.📚 Better learning and onboarding
🤖 Improved docs assistant experience
Overall, 8.3.234 is a safe, drop-in update: you can upgrade to benefit from better security, stability, and docs UX with no changes to your existing YOLO training or inference code. ✅
onnxruntime-gpu download from Ultralytics assets by @lakshanthad in https://github.com/ultralytics/ultralytics/pull/22823ultralytics 8.3.234 Security fix: ast.literal_eval for safer evaluation of metadata strings by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/22847Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.233...v8.3.234
Ultralytics 8.3.233 makes Sony IMX model export more robust across Debian/Ubuntu/Raspberry Pi systems, adds a coming‑soon Axelera edge‑AI integration
Ultralytics 8.3.233 makes Sony IMX model export more robust across Debian/Ubuntu/Raspberry Pi systems, adds a coming‑soon Axelera edge‑AI integration doc, and improves Jetson + docs tooling for smoother deployments and documentation. 🚀
🧩 Sony IMX exporter: Debian & Java handling improved (main change)
ultralytics.utils:
is_debian() helper.IS_DEBIAN, IS_DEBIAN_BOOKWORM, IS_DEBIAN_TRIXIE, plus IS_UBUNTU and existing IS_RASPBERRYPI.export_imx now installs the correct Java version per platform:
sudo when available, and logs what it’s doing for clarity.🧠 New Axelera integration docs (coming soon)
docs/en/integrations/axelera.md page and linked it in the integrations index and mkdocs.yml.🤖 NVIDIA Jetson guide updates
onnxruntime-gpu 1.23.0 on JetPack 6 + Python 3.10 via Jetson AI Lab wheel.onnxruntime-gpu 1.20.0 example as a fallback.🔗 Axelera doc link clean‑up
www.ultralytics.com URLs:
🧹 Quieter, cleaner docs builds
docs/build_docs.py now uses:
git clone -q --depth=1 --single-branch -b main ...hub-sdk and docs repos.🔍 Utils reference docs extended
ultralytics.utils.__init__.is_debian to the reference docs.🏷️ Version bump
8.3.233.✅ More reliable Sony IMX deployments on edge Linux systems
🌍 Better cross‑platform support story
IS_DEBIAN_BOOKWORM, IS_DEBIAN_TRIXIE, IS_UBUNTU, IS_RASPBERRYPI make it easier to:
🚀 Future‑proof edge inference with Axelera
format="axelera" once runtime wheels are available.⚙️ Smoother NVIDIA Jetson experience
onnxruntime-gpu instructions for JetPack 6 reduce install friction and compatibility issues when exporting certain formats.📚 Cleaner docs & navigation
Overall, 8.3.233 is a stability and platform‑support focused release, especially valuable if you’re exporting to Sony IMX, working on Debian/Ubuntu/Raspberry Pi, planning Axelera edge deployments, or using NVIDIA Jetson devices. 🌐💡
ultralytics 8.3.233 Add Debian OS + Java install support to Sony IMX exporter by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/22814Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.232...v8.3.233
Ultralytics 8.3.232 is a small but important release that fixes OBB export behavior for more reliable deployments, improves ARM64 Docker images, and s
Ultralytics 8.3.232 is a small but important release that fixes OBB export behavior for more reliable deployments, improves ARM64 Docker images, and significantly polishes documentation UX, search, and SAM 3 docs. 🚀
🚧 OBB export fix (core model behavior)
ExporterDetectionModel.forward.8 for OBB NMS instead of scaling by the number of classes.🐧 Docker: ARM64 image modernized
arm64v8/debian:bookworm-slim to arm64v8/ubuntu:24.04.apt with apt upgrade -y for up-to-date, more secure images.🧮 GPU/cuDNN compatibility helper
yolo-common-issues.md to check:
📚 SAM 3 documentation aligned with Meta’s release
pip install ultralytics.🔍 Docs search experience upgrade
🧭 Docs navigation & links cleanup
.md files; converts them to trailing-slash internal paths.models/yolov5.md).📊 Mermaid diagrams in docs
pymdownx.superfences configuration in mkdocs.yml.```mermaid blocks.📘 Docs build & UI tweaks
--strict from zensical build to reduce build failures on minor warnings.mkdocs_github_authors.yaml.More reliable OBB deployments 🧱
Smoother cross-platform usage & CI 🧪
--strict) reduces CI friction and local build failures.Simpler troubleshooting for GPU issues 🔧
Better documentation experience 📖
Overall, 8.3.232 is a stability and UX-focused release: production users gain a critical OBB export fix, while everyone benefits from better docs, search, and environment support. ✨
ultralytics 8.3.232 Fix OBB NMS export lower box count by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22783Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.231...v8.3.232
✅ No breaking changes to YOLO11 models
v8.3.231 is a docs and tooling–focused release 🚀 that introduces a new Zensical-powered documentation pipeline with rich API reference pages, better HTML post-processing, and multiple reliability and usability improvements—without changing core YOLO11 model behavior.
📚 New Zensical documentation engine (major)
mkdocs build to zensical build with a revamped docs/build_docs.py.build_reference_docs pipeline that:
mkdocs-ultralytics-plugin for better metadata, JSON-LD, authors, and social cards.🧩 MiniJinja-based docs macros (templating refresh)
render_jinja_macros).mkdocs.yml extra values,ultralytics/cfg/default.yaml,indent filter to stay compatible with existing macros.docs/en, docs/macros) are backed up and restored around the build to keep the repo clean.🧱 Reference docs re-architecture
docs/build_reference.py rewritten to:
__init__ docstring info into class docs intelligently (no runtime behavior changes).🎨 Docs UX & content polish
/cuda-zone → /cuda.doc.dvc.org domain.CASIA-LMC-Lab/FastSAM repository.!!! note, !!! warning) so they render reliably.--8<-- "docs/macros/..." includes, consistent with the new macro pipeline.⚙️ Settings & concurrency reliability
SettingsManager._load to call super().update(...) instead of self.update(...).🍎 Apple MPS & metrics cleanup
OKS_SIGMA to float32 in metrics for more consistent pose evaluation behavior.🖼️ Data format documentation clarification
uint8, in addition to (channel, height, width) layout and .tiff/.tif extensions. Prevents subtle runtime issues when preparing multispectral data. 🌈💻 Windows workflow documentation
train.md) and validation (val.md) docs now:
if __name__ == "__main__":
...
🧹 Docstring cleanup for codebase
Examples: and Attributes: blocks from many class __init__ docstrings.🧾 Tooling & repo hygiene
mkdocs-ultralytics-plugin dev dependency to >=0.2.2 to support the new docs flow.uv.lock to .gitignore to avoid committing local UV lock files.JUPYTER_PLATFORM_DIRS export from the docs CI workflow.🧠 Richer, easier-to-navigate API docs
🛠️ More robust and maintainable docs pipeline
🔒 Safer multi-process / multi-instance usage
SettingsManager read behavior reduces the chance of config corruption or race conditions when multiple processes touch settings simultaneously.💻 Smoother platform-specific experiences
🛰️ Fewer user mistakes with advanced data / integrations
uint8 guidance prevents subtle multispectral dataset bugs.✅ No breaking changes to YOLO11 models
uint8 by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22756SettingsManager to avoid race conditions in parallel instances by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22762ultralytics 8.3.231 New Zensical docs refresh by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22761Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.230...v8.3.231
ultralytics 8.3.230 is a stability and UX-focused release that fixes color issues in visualizations, improves TensorRT and device handling, and loosen
ultralytics 8.3.230 is a stability and UX-focused release that fixes color issues in visualizations, improves TensorRT and device handling, and loosens dependency constraints for smoother installs, along with some docs and CI polish. 🎨⚙️🚀
🎨 Accurate colors for visualizations (Tensor + PIL) [@Y-T-G, #22734]
pil=True plots return proper RGB images instead of visually “off” colors.Results.plot() now delegates to annotator.result(pil) for consistent behavior.⚙️ More robust TensorRT dynamic input handling [@Laughing-q, #22729]
🧠 Safer device selection for CPU / MPS [@Y-T-G, #22740]
select_device("cpu") or select_device("mps"), CUDA_VISIBLE_DEVICES is now set to "" instead of "-1".📦 Looser dependency version pins for better compatibility [@glenn-jocher, #22721]
numpy, matplotlib, opencv-python, pillow, pyyaml, requests, scipy, torchvision, psutil, polars.torch>=1.8.0torch>=1.8.0,!=2.4.0 (skips known-bad 2.4.0 on Windows CPU).ultralytics-thop now uses >=2.0.18 instead of an upper cap.🖼️ Docs UX + SEO improvements [@sergiuwaxmann, @glenn-jocher, @RizwanMunawar]
page metadata, preventing template errors.J1BaCqytBmA), aligned with current YOLO11 workflows.🧪 CI & workflow maintenance [@dependabot[bot]]
actions/checkout@v6 (including CI, docs, docker, link checks, publishing, mirroring).🎨 More trustworthy visual outputs
pil=True for plotting or UI integrations.🧩 Smoother deployment on TensorRT and varied input sizes
💻 More reliable CPU/MPS runs and Ultralytics HUB integration
CUDA_VISIBLE_DEVICES properly avoids crashes and confusing CUDA behavior on systems where GPUs must be “hidden.”📦 Easier installs and upgrades
numpy, torch, torchvision).📚 Better docs experience & onboarding
🏗️ More robust CI/CD
actions/checkout@v6 keeps the project’s automation current, aiding long-term reliability of builds, tests, and releases.Overall, v8.3.230 is a quality-focused release: if you visualize results, run dynamic TensorRT models, rely on CPU/MPS, or manage modern Python stacks, you should see more stable and predictable behavior. ✅
dynamic TensorRT model by setting the max_shape as input_shape by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/22729CUDA_VISIBLE_DEVICES to empty string instead of -1 to avoid Windows crash by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22740ultralytics 8.3.230 Fix color conversion for Tensor inputs and PIL plots by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22734Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.229...v8.3.230
Dependency upper bounds in pyproject.toml (PR #22701) to avoid breaking changes in numpy, torch, opencv, pillow, polars, etc.
YOLO 8.3.229 focuses on much faster COCO instance segmentation validation (~3×), tighter dependency and environment controls, improved multi‑GPU behavior, and a broad documentation refresh (Jetson, datasets, guides) for smoother real‑world use. 🚀🧠
⚡ 300% Faster COCO Segmentation Validation (PR #22651)
faster_coco_eval RLE encoding with a native, optimized RLE pipeline.ops.scale_masks API with ratio_pad support.🔐 Safer, More Predictable Environments
pyproject.toml (PR #22701) to avoid breaking changes in numpy, torch, opencv, pillow, polars, etc.torch constraints to avoid known Windows issues.python=3.12 and torch>=2.9 with clear errors.uv from the base image (PR #22709).🖥️ More Robust Multi‑GPU & CI
coco128 training for better multi‑GPU validation (PR #22710).batch_size >= dataset_size (PR #22714): safely falls back to batch_size=1 to avoid failures on tiny datasets.2.329.0 (PR #22707).📚 Major Documentation & Guide Updates
📖 Docs Build & Chat Widget Improvements
script, pre, code, textarea content and regex‑based HTML cleanups (PRs #22676, #22678).chat.js v0.0.6 with separate chatExamples and searchExamples for better in‑docs help (PRs #22680, #22691, #22696).🎨 Repo & Config Quality
print-width updated to 120 and applied across docs JS/HTML/CSS plus dataset YAML for cleaner diffs and readability (PR #22706).🎥 & Learning Experience
Faster Experiment Cycles for Segmentation
More Stable, Reproducible Installs
More Reliable Multi‑GPU Training & Testing
Better Docs for Real Deployments
Smoother Documentation UX
Overall, v8.3.229 is a performance‑ and robustness‑oriented release: your COCO segmentation validation is faster, environments are safer, multi‑GPU workflows are more predictable, and the documentation better reflects how to run YOLO11 in real systems. 🌐🔥
script tags in JS minify by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22678python>=3.8<3.12 and torch<2.9 for IMX export by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22673ultralytics/llm@v0.0.5 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22691print-width 120 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22706batch_size is greater than the total dataset size by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22714pip install uv from Dockerfile by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22709gpu-latest runners in ci.yml by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22710ultralytics 8.3.229 300% faster COCO Segmentation val by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/22651Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.228...v8.3.229
CLIP/MobileCLIP tokenization is now safer by default with a new truncate option, preventing crashes on long prompts, alongside improved progress-bar r
CLIP/MobileCLIP tokenization is now safer by default with a new truncate option, preventing crashes on long prompts, alongside improved progress-bar readability, better docs search/chat UX, augmentation docs enhancements, and more reliable ARM CI. 🚀
CLIP/MobileCLIP safety upgrade (priority)
truncate: bool = True to CLIP.tokenize and MobileCLIPTS.tokenize to handle overlength text gracefully.truncate=False for debugging or exact validation.Progress bar readability
1.5s/it instead of 0.0it/s). Thanks @fcakyon (PR #22660).Docs chat/search experience
Data augmentation docs polish
augmentations argument for custom Albumentations pipelines to the macro table (PR #22615).hsv_s, not hsv_h) for accuracy (PR #22645).CI reliability
More robust text models ✨
truncate=False) for developers who need exact-length checks or debugging.Clearer training feedback 📈
Better docs experience 🔎
More flexible augmentations 🧪
augmentations unlocks custom Albumentations transforms via Python API, enabling advanced experimentation across detect/segment/pose/obb tasks.Improved contributor reliability 🧹
Example usage (strict vs safe tokenization):
from ultralytics.nn.text_model import CLIP, MobileCLIPTS
clip_model = CLIP(size="ViT-B/32", device="cpu")
safe_tokens = clip_model.tokenize("a very long caption ...", truncate=True) # default, avoids errors
strict_tokens = clip_model.tokenize("a very long caption ...", truncate=False) # may raise if too long
mobileclip = MobileCLIPTS(device="cpu")
tokens = mobileclip.tokenize(["caption 1", "caption 2"]) # safely truncated by default
chat.js bot by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22639ultralytics 8.3.228 Fix CLIP token truncation by @h13-0 in https://github.com/ultralytics/ultralytics/pull/22650Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.227...v8.3.228
Faster patching of vulnerabilities and tooling updates via daily Dependabot runs. ⚡
Stability-focused release: pins ONNX to avoid export breakages, improves Edge TPU export setup on modern Linux, and includes several quality-of-life updates across docs, dependencies, and tooling. ✅🛠️
onnx>=1.12.0,<=1.19.1 to ensure reliable ONNX and TensorFlow SavedModel exports, especially in Conda environments. (See PR #22628) 🔒scipy-stubs>=1.14.1.4 for Python 3.10+. (See PR #22634) 🧠.gitignore to prevent large media commits. (See PR #22635) 📁🚫export and smoother CI. ✅Quick tip: If you hit ONNX-related export issues, ensure ONNX is pinned locally:
pip install "onnx<=1.19.1"
scipy-stubs dependency to require version >=1.14.1.4 for Python 3.10+ by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/22634citation of african-wildlife dataset by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/22632ultralytics 8.3.227 Pin onnx<=1.19.1 for Conda compatibility by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/22628Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.226...v8.3.227
Security-first release that replaces unsafe eval() with safe parsing across the codebase, alongside quality-of-life upgrades: custom Albumentations in
Security-first release that replaces unsafe eval() with safe parsing across the codebase, alongside quality-of-life upgrades: custom Albumentations in Python API, simplified NCNN export via PNNX, multi-GPU training stability fixes, and improved installer/logging and docs. 🔒⚙️
Security & Parsing (priority)
Data Augmentation
augmentations parameter in model.train(...). 🎨Export/Deployment
Training & Stability
val=False by ensuring final_epoch is always defined; added CUDA test coverage. 🧪Installation & Tooling
check_requirements() logs and error handling, especially with the uv package manager (stderr merged into stdout; better fallback behavior). 📦Docs & UX
Safer by default
Better training experience
Smoother export and deployment
More reliable tooling and docs
Example: Using custom Albumentations in Python API
import albumentations as A
from ultralytics import YOLO
model = YOLO("yolo11n.pt")
custom_transforms = [
A.Blur(blur_limit=7, p=0.5),
A.CLAHE(clip_limit=4.0, p=0.5),
]
model.train(data="coco8.yaml", epochs=100, imgsz=640, augmentations=custom_transforms)
Helpful links:
KITTI notebook in docs by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/22592final_epoch for DDP training when val=False by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22607check_requirements() missing output when using uv package manager by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22623ultralytics 8.3.226 ♻️ Replace eval() with ast.literal_eval() for security by @onuralpszr in https://github.com/ultralytics/ultralytics/pull/22597Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.225...v8.3.226
SAM-2 interactive predictor now uses 0-based object IDs (breaking change) 🔁
Ultralytics 8.3.225 adds first-class KITTI dataset support for YOLO11, improves model checkpoint loading reliability, cleans up docs/tests, and makes Jetson builds more reproducible. 🚗🧠⚙️
New: KITTI dataset support and docs 🎉
kitti.yaml with classes and download link.Safer checkpoint loading for SAM/SAM2 🛡️
torch_load helper used for SAM and SAM2 checkpoint loading for better cross-version PyTorch compatibility.SAM-2 interactive predictor now uses 0-based object IDs (breaking change) 🔁
Jetson JetPack 4 Docker build reliability 🔧
Docs polish and CI stability 📚✅
Faster start on autonomous driving research 🚦
More reliable model loading across environments 🔄
Clearer APIs and fewer indexing mistakes 🧭
Reproducible hardware builds 🧪
Better developer and user experience ✨
Quick start on KITTI with YOLO11:
yolo detect train data=kitti.yaml model=yolo11n.pt epochs=100 imgsz=640
from ultralytics import YOLO
model = YOLO("yolo11n.pt")
model.train(data="kitti.yaml", epochs=100, imgsz=640)
torch_load patch for weights_only cases by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/22577integrations/executorch.md examples by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/22581Executorch test by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/22582ultralytics 8.3.225 New KITTI dataset by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/22539Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.224...v8.3.225
Ultralytics 8.3.224 boosts segmentation performance and stability on GPUs, adds CPU-only ExecuTorch export/benchmarks for YOLO11, fixes multi‑GPU eval
Ultralytics 8.3.224 boosts segmentation performance and stability on GPUs, adds CPU-only ExecuTorch export/benchmarks for YOLO11, fixes multi‑GPU evaluation and ONNX device selection, and improves logging + docs for a smoother developer experience. 🚀
boxes live on the same device as masks and avoids slow Python loops on CUDA.dist.gather_object and cleans up worker memory.device_id to CUDAExecutionProvider to avoid accidental GPU 0 usage.download-binaries for ONNX Runtime to reduce setup friction and version conflicts.cuda:1) is honored, improving reliability in multi-GPU environments.Quick tips:
yolo export model=yolo11n.pt format=executorch device=cpu
jdict for COCO evaluation by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/22541ultralytics 8.3.224 Accelerate crop_mask on GPU with is_cuda check by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/22575Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.223...v8.3.224
Ultralytics 8.3.223 improves post-processing correctness in FastNMS, stabilizes RT-DETR validation, and refactors TensorFlow/TFLite/Edge TPU/TFJS expo
Ultralytics 8.3.223 improves post-processing correctness in FastNMS, stabilizes RT-DETR validation, and refactors TensorFlow/TFLite/Edge TPU/TFJS export for cleaner APIs and more reliable exports—plus better FLOPs accuracy via a dependency bump. 🚀
FastNMS fix (priority) 🔧
scores in-place when use_triu=False, ensuring alignment with sorted indices for consistent outputs in downstream consumers like NMSModel.RT-DETR validation stability ✅
ratio_pad is correctly formatted during validation to prevent crashes and stabilize evaluation.Export refactor: TensorFlow/TFLite/Edge TPU/TFJS ♻️
ultralytics.utils.export.engine and ultralytics.utils.export.tensorflow with wrappers to avoid fragile ops and improve reliability.Accurate FLOPs reporting 📈
ultralytics-thop>=2.0.18 for better FLOPs calculations, improving benchmarking across models (e.g., YOLO11/YOLO26).IMX export message clarity 🧭
Standardized dataset class names 🇺🇸
More correct and stable NMS outputs 🛡️
scores, preventing subtle ranking issues and improving result consistency for all users relying on FastNMS.Fewer validation crashes for RT-DETR 📏
ratio_pad handling avoids runtime errors and increases reliability during validation and metrics computation.Easier, more reliable exports across ecosystems 🌐
Better, more consistent benchmarks 🔬
ultralytics-thop improves FLOPs accuracy and compatibility, leading to more trustworthy comparisons across models and environments.Clearer developer experience ✨
Quick tip to upgrade:
pip install -U ultralytics
Thanks to contributors @Y-T-G, @Laughing-q, @glenn-jocher, and @lakshanthad! 🙌
ultralytics-thop>=2.0.18 by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22526saved_model/pb/edgetpu/tfjs formats by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/22115ratio_pad in RTDETRDataset by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22525ultralytics 8.3.223 Sort scores inplace in FastNMS with use_triu=False by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22537Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.222...v8.3.223
TFLite INT8 export is now enabled for end2end models, plus improved quantized NMS accuracy and full IMX500 Classification support—expanding fast, smal
TFLite INT8 export is now enabled for end2end models, plus improved quantized NMS accuracy and full IMX500 Classification support—expanding fast, small-footprint deployments across mobile and edge devices. 🚀📱
model.end2end=True (Enable TFLite INT8 export for end2end models – PR #22503).yolo export model=yolov10n.pt format=tflite int8yolo val task=detect model=yolov10n_saved_model/yolov10n_int8.tflite data=coco128.yamlHappy exporting and deploying! ⚡
RUF and FA ruff formatting by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22480res to results by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/22501ultralytics 8.3.222 Enable TFLite INT8 export for end2end models by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22503Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.221...v8.3.222
v8.3.221 focuses on a large Ruff-driven codebase refactor that modernizes typing, cleans imports/variables, and polishes formatting/logging with minim
v8.3.221 focuses on a large Ruff-driven codebase refactor that modernizes typing, cleans imports/variables, and polishes formatting/logging with minimal behavioral change. It also tightens docs CI linting, fixes exception logging patterns, updates ExecuTorch docs links, and improves test isolation. ✨
Path | None, dict | None).# noqa, standardized imports, safer disk-usage vars.next(...) for first matches, clearer f-string formatting with !s, !r).int(round(...)) → round(...).__all__ exports and minor logging/output polish.str(e) to f"{e}" across code and docs for cleaner messages.tmp_path for better isolation and parallel test stability, replacing ad-hoc TMP dir usage.References:
tmp_path improves CI stability and confidence in releases.Upgrade note: No action required. If contributing, ensure your PRs pass the stricter Ruff checks and follow the updated exception/logging patterns. ✅
f"{e}" by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22473--extend-select F in docs.yml by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22474tmp_path for improved test isolation by @Borda in https://github.com/ultralytics/ultralytics/pull/22456ultralytics 8.3.221 Ruff RUF refactor by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22475Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.220...v8.3.221
ExecuTorch export lands! You can now export YOLO11/YOLO26 models to .pte with XNNPACK acceleration for fast, native on-device inference on mobile and
ExecuTorch export lands! You can now export YOLO11/YOLO26 models to .pte with XNNPACK acceleration for fast, native on-device inference on mobile and edge devices. Plus, a stability fix for CLIP embedding extraction in VisualAI. 📱⚡🧠
format=executorch → outputs a .pte model inside a <name>_executorch_model/ directoryexport_formats() and the exporter flowexport_executorch() method using torch.export, XnnpackPartitioner, and to_edge_transform_and_lower.pte loading and inference dispatchMinimal usage:
yolo export model=yolo11n.pt format=executorchfrom ultralytics import YOLO
YOLO("yolo11n.pt").export(format="executorch")
Helpful links:
.pte for running models on iOS, Android, and embedded Linux devices via ExecuTorchexecutorch and flatbuffers installed.pte and metadata.yaml; inference via ExecuTorch runtime (not standard Python YOLO() loading)Overall, this release expands deployment options for YOLO11/YOLO26—especially for mobile and edge users—while improving robustness in VisualAI workflows. 🎉
VisualAI clip feature extraction numpy() grad error by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/22463ultralytics 8.3.220 Add ExecuTorch export (.pte) with XNNPACK by @ambitious-octopus in https://github.com/ultralytics/ultralytics/pull/22244Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.219...v8.3.220
FastSAM gets a reliable mask-resizing fix and a cleaner, built-in CLIP integration for prompt-based segmentation—improving accuracy, stability, and ea
FastSAM gets a reliable mask-resizing fix and a cleaner, built-in CLIP integration for prompt-based segmentation—improving accuracy, stability, and ease of use. 🚀🖼️
ultralytics.nn.text_model.CLIP, standardizing image/text feature extraction and similarity scoring.8.3.219.Reference: See the current PR FastSAM masks .float() upsampling fix (PR #22460) by @Y-T-G.
ultralytics 8.3.219 FastSAM masks .float() upsampling fix by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22460Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.218...v8.3.219
Better, faster multi-GPU training: v8.3.218 enables true multi-GPU validation during training with correct cross-GPU metric aggregation and a new cont
Better, faster multi-GPU training: v8.3.218 enables true multi-GPU validation during training with correct cross-GPU metric aggregation and a new contiguous sampler for stable evaluation. 🚀
shuffle=False (e.g., rect/size-grouped evaluation) to prevent interleaved indices.DistributedSampler when shuffle=True.ContiguousDistributedSampler.Links:
rect=True.Quick tip to run distributed training and benefit from these improvements:
yolo detect train data=coco128.yaml model=yolo11n.pt devices=0,1,2,3from ultralytics import YOLO
model = YOLO("yolo11n.pt")
model.train(data="coco128.yaml", devices=[0, 1], imgsz=640, epochs=50)
Happy training and validating across GPUs! 🎉
ultralytics 8.3.218 Enable multi-GPU validation during training by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22377Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.217...v8.3.218
Segmentation gets leaner and more reliable: segment masks are now ~4× lighter with consistent uint8 handling, plus smoother first-iteration latency th
Segmentation gets leaner and more reliable: segment masks are now ~4× lighter with consistent uint8 handling, plus smoother first-iteration latency thanks to NMS warmup and a new dataloader pin_memory option. 🚀🧠
Segmentation masks optimized and standardized (PR: Segment masks now 4× lighter with .byte() by @glenn-jocher)
.byte() (uint8) across processing and plotting, reducing memory and avoiding dtype mismatches.process_mask and process_mask_native return uint8 masks instead of bool.masks2segments consumes byte masks directly (no extra cast).float() before IoU to prevent edge-case errors..byte() to ensure consistent uint8 NumPy output..byte() optimization.Dataloader and backend improvements (PR: Add NMS warmup for clearer post-processing latency by @Y-T-G)
pin_memory parameter in build_dataloader(..., pin_memory: bool = True).pin_memory=self.training, reducing host memory pinning during eval by default.Version bump to 8.3.217.
Faster, lighter segmentation workflows 💾⚡
More robust and consistent results ✅
Smoother first-iteration performance 🚀
Better memory control for training/eval 🧰
pin_memory flag allows fine-grained control to balance throughput and system memory behavior; disabled by default in validation for stability.Minimal examples:
from ultralytics.data.build import build_dataloader
# dataset = ...
dl = build_dataloader(dataset, batch=16, workers=8, pin_memory=False)
Contributors: @glenn-jocher, @Y-T-G, @Laughing-q
Links:
.byte() optimization (PR #22427)ultralytics 8.3.217 Segment masks now 4× lighter with .byte() optimization by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22427Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.216...v8.3.217
Reduces unexpected keyword errors and deprecation warnings across environments.
Ultralytics 8.3.216 speeds up and stabilizes segmentation mask rendering, improves pose keypoint metadata handling, and enhances checkpoint loading compatibility across PyTorch versions. ⚡🖼️🧩
Faster and more reliable mask plotting in the Annotator and Plotter
retina_masks=False using ops.scale_masks.ops.process_mask with broadcasted ratios and named arguments.Pose models: keypoint names metadata (kpt_names) support end-to-end
model.kpt_names; exporter saves kpt_names into model metadata.kpt_names for downstream use and exports.More compatible checkpoint loading for SAM/SAM 2
torch.load(..., weights_only=False) when supported, with safe fallback for older PyTorch.Documentation enhancement
Faster, cleaner segmentation overlays 🎨
Better pose model usability 🕺
from ultralytics import YOLO
model = YOLO("yolo11n-pose.pt")
print(model.kpt_names)
More robust model loading 🧠
Improved learning experience 📚
Version bump: 8.3.215 → 8.3.216. 🚀
weights_only parameter by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22421model-yaml-config.md by @RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/22414ultralytics 8.3.216 Faster Annotator mask plotting by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22419Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.215...v8.3.216
Faster, more stable instance segmentation and more reliable exports. This release accelerates mask cropping by ~3x, hardens ONNX/RT-DETR exports, and
Faster, more stable instance segmentation and more reliable exports. This release accelerates mask cropping by ~3x, hardens ONNX/RT-DETR exports, and restores official Jetson JetPack 5 Docker builds — all while refreshing docs and tutorials. ⚡️🛡️🐧
✂️ Segmentation speedup (priority)
utils.ops.crop_mask for ~3x faster mask cropping on small batches (e.g., 6 masks on M4 MacBook Pro) by using efficient per-mask slicing for n<50 and vectorized logic for larger batches.amax-based check) to avoid slow indexing on common cases.🧭 ONNX export reliability
dynamo=False by default on PyTorch 2.4 to avoid Dynamo-related instability during export; transparent to users.🚫 RT-DETR export safeguards
RTDETRDecoder) and forces nms=False for end-to-end/RT-DETR exports.📦 Export Docker stability
torch<=2.8.0 in the Python export Docker image to avoid IMX issues with 2.9.0; keeps numpy==1.26.4.🐧 Jetson support
📚 Docs and examples
dir='datasets'.Tip: Upgrade with pip install -U ultralytics and enjoy speedups automatically—no code changes needed for segmentation or export workflows.
dynamo=False by default for ONNX exports by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22388torch<=2.8.0 in Dockerfile-python-export by @Laughing-q in https://github.com/ultralytics/ultralytics/pull/22392nms export for RTDETR models by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22385ultralytics 8.3.215 Segmentation crop_masks speedup by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22386Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.214...v8.3.215
Cleaner experiment logs and more robust training—no breaking changes. 🚀
Adds confidence scores to classification validation plots and strengthens NaN recovery during training, plus documentation updates highlighting SAM 3 and YOLO26. Cleaner experiment logs and more robust training—no breaking changes. 🚀
self.loss (not self.tloss) and refine NaN/fitness collapse detection logic for more reliable recovery. See Improve NaN handler loss check (PR #22382).results.csv when not resuming so new runs don’t append to old logs with exist_ok=True. See Reset results.csv behavior (PR #22364).results.csv per run (unless resuming) keeps metrics tidy and comparable. 🧹Tip: Upgrade with pip install -U ultralytics to get these improvements. ✨
results.csv if not resuming by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22364InferenceTensor error during NaN recovery by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22372self.loss instead of self.tloss by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22382ultralytics 8.3.214 Show confidence in classification plots by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22365Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.213...v8.3.214
A stability-focused release that automatically recovers from NaN/Inf training issues and cleans up resume logic, plus faster Objects365 setup and unif
A stability-focused release that automatically recovers from NaN/Inf training issues and cleans up resume logic, plus faster Objects365 setup and unified dataset download URLs. 🚑🔁⚡
Training robustness and resume improvements (primary change)
_load_checkpoint_state() now used by resume_training() to reduce duplication and state drift.test_nan_recovery injects a NaN to verify recovery path. ✅Dataset YAMLs: unified asset URLs
ASSETS_URL across VOC, COCO, COCO-Pose, VisDrone, and LVIS for maintainable, consistent downloads. 🔗Objects365 setup speedups
ThreadPoolExecutor.CI and test updates
Version bump
ultralytics now 8.3.213.More resilient training at scale
Faster dataset preparation
More reliable downloads
Clearer platform expectations
Quick start
pip install -U ultralyticslast.pt and continue.ASSETS_URL in dataset YAMLs by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22361ultralytics 8.3.213 NaN epoch recovery by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22352Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.212...v8.3.213
Ultralytics 8.3.212 focuses on making training more robust and predictable by hardening the Trainer against edge cases (like deleted save directories
Ultralytics 8.3.212 focuses on making training more robust and predictable by hardening the Trainer against edge cases (like deleted save directories and transient non-finite losses), while simplifying the optimizer step for modern PyTorch. 🚀
read_results_csv() now returns {} on read failures instead of raising. 📄➡️{}best.pt, last.pt, results.csv). 💾🗂️clip_grad_norm_ with max_norm=10.0), scaler.step(), scaler.update(), zero grads. 🧹uv for faster, deterministic installs. ⚡🐳save_dir is deleted mid-run or on flaky filesystems.results.csv; failures return {} so training and tools can continue gracefully.uv, improving development velocity without affecting end users.Tip: If you programmatically consume training metrics, you can safely handle missing/locked CSVs:
from ultralytics.engine.trainer import BaseTrainer
trainer = BaseTrainer(args={})
metrics = trainer.read_results_csv() # returns {} on failure instead of raising
uv for docker.yml pip install tests by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22356ultralytics 8.3.212 Improve Trainer robustness to save_dir deletion by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22358Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.211...v8.3.212
More robust training and downloads. This release skips optimizer updates on non‑finite gradients (NaN/Inf) to prevent crashes and centralizes asset UR
More robust training and downloads. This release skips optimizer updates on non‑finite gradients (NaN/Inf) to prevent crashes and centralizes asset URLs for more reliable tests and dataset downloads. 🛡️🌐
optimizer_step(): gradient clipping with non‑finite detection and safe skip of the optimizer update instead of crashing.ASSETS_URL constant and replaced hardcoded asset links across tests, data utils, benchmarks, and converters.downloads.is_url() is now safer and faster:
safe_download() now aliases ASSETS_URL to the public assets host automatically, simplifying host switching.8.3.211.Key PRs:
ASSETS_URL and improve download robustness — see PR Use ASSETS_URL in testsASSETS_URL reduces flaky tests and download issues across environments.Quick start or upgrade:
pip install -U ultralyticsTip: If you see warnings about non‑finite loss or gradients, try:
amp=False).ASSETS_URL in tests by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22355ultralytics 8.3.211 Skip non-finite gradients in optimizer_step() by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22350Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.210...v8.3.211
More stable training by skipping non-finite loss batches, smoother RT-DETR device/precision moves, and better ONNX profiling for multi-input/dynamic-s
More stable training by skipping non-finite loss batches, smoother RT-DETR device/precision moves, and better ONNX profiling for multi-input/dynamic-shape models. 🚀
Training stability
RT-DETR device/precision consistency
_apply to RTDETRDetectionModel so anchors and masks follow .to(), .cuda(), .cpu(), .half(), etc. 🔁 (See PR: Add _apply to RTDETR)ONNX benchmarking improvements
More reliable training out of the box
Fewer device/dtype gotchas for RT-DETR
Robust and accurate ONNX profiling
Quick update tip:
pip install -U ultralytics ✅_apply method to RTDETR for better device consistency by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22341ultralytics 8.3.210 Skip non-finite training forward passes by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22353Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.209...v8.3.210
RT-DETR exports to TensorFlow/TFLite are now reliable by automatically using ONNX opset 19, plus a small tuner bug fix and documentation/CI updates. ✅
RT-DETR exports to TensorFlow/TFLite are now reliable by automatically using ONNX opset 19, plus a small tuner bug fix and documentation/CI updates. ✅🚀
RT-DETR TensorFlow/TFLite export made robust
opset=19 during the ONNX step when exporting RT-DETR to TensorFlow SavedModel (and downstream TFLite). See the PR by @Y-T-G: RTDETR TFLite export fix with ONNX opset constraints fixing issue #18055. 🔧simplify=True), improving export stability.Dependency updates for export
onnxslim to >=0.1.71 for both ONNX and TensorFlow SavedModel export paths for better model graph simplification. 📦Tuner reliability fix
CI and Docs improvements
More reliable RT-DETR → TensorFlow/TFLite exports
Smoother export experience
onnxslim enhances ONNX graph simplification, reducing export friction. ⏩More accurate tuning results
Clearer docs and steadier CI
Quick examples:
yolo export model=rtdetr.pt format=tf or yolo export model=rtdetr.pt format=tflitefrom ultralytics import YOLO
model = YOLO("rtdetr.pt")
model.export(format="tf") # TensorFlow SavedModel
model.export(format="tflite") # TFLite
ultralytics/stars to Docs CI page by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22331ultralytics 8.3.209 RTDETR TFLite export fix with ONNX opset constraints by @Y-T-G in https://github.com/ultralytics/ultralytics/pull/22314Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.208...v8.3.209
Tidy Dependabot config and labels; remove deprecated reviewer field (PR #22322). 🧹
Expands safe FP16 (half-precision) export support across ONNX and TorchScript, improves NCNN export reliability by switching to the ONNX pipeline, and delivers faster RT-DETR inference via anchor caching — plus clearer docs and smoother dependency handling. 🚀
FP16 export behavior (PR #22316)
half=True on CPU, it now warns and automatically uses half=False.NCNN export pipeline update (PR #22315)
format='ncnn'. TorchScript is only produced if explicitly requested. 🔄RT-DETR performance optimization (PR #22318)
dynamic=True.Smarter dependency checks with interchangeable packages (PR #22321)
check_requirements() supports alternatives like ("onnxruntime", "onnxruntime-gpu"), making installs more flexible. ✅Validation docs clarity (PR #22319)
model.names (the model’s class set), not the dataset YAML classes. 📝Better examples for check_requirements usage (PR #22317)
CI and maintenance improvements
Safer, clearer FP16 exports on GPU only
More reliable NCNN exports
Faster RT-DETR workflows
Smoother installs and environment setup
onnxruntime vs onnxruntime-gpu) reduces friction across CPU/GPU setups.Reduced test flakes and cleaner maintenance
Quick examples:
from ultralytics import YOLO
YOLO("yolo11n.pt").export(format="onnx", half=True, device=0) # ✅ GPU FP16
YOLO("yolo11n.pt").export(format="onnx", half=True, device="cpu") # ℹ️ Will warn and use half=False
from ultralytics.utils.checks import check_requirements
check_requirements([("onnxruntime", "onnxruntime-gpu"), "numpy"])
check_requirements pip install extras example by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22317missing_ok=True to TensorRT INT8 cache cleanup test by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22320check_requirements() by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22321ultralytics 8.3.208 Expand FP16 export support by @glenn-jocher in https://github.com/ultralytics/ultralytics/pull/22316Full Changelog: https://github.com/ultralytics/ultralytics/compare/v8.3.207...v8.3.208
One column per quarter.