Ultralytics v8.4.214
v8.4.214 applies EXIF orientation consistently when reading images across formats, read flags, and OpenCV versions.
The Deploy Log | Products | Free, no signup
Ultralytics changelog
26 versions of Ultralytics from Ultralytics in The Deploy Log. From to . On the median a new version every 1.5 days, publisher's dates only; the longest wait was 8 days. The newest is v8.4.214. 23 versions in the last 30 days. Each links the publisher's own page.
| Version | Publisher's date | Days after the one before |
|---|---|---|
| v8.4.214 | 1 | |
| 8.4.213 | 1 | |
| 8.4.212 | 2 | |
| v8.4.211 | 1 | |
| 8.4.210 | 2 | |
| 8.4.208 | 1 | |
| 8.4.207 | 0 | |
| 8.4.206 | 1 | |
| 8.4.204 | 1 | |
| v8.4.203 | 2 | |
| v8.4.201 | 0 | |
| 8.4.200 | 0 | |
| 8.4.199 | 4 | |
| 8.4.194 | 0 | |
| 8.4.193 | 2 | |
| v8.4.192 | 1 | |
| 8.4.191 | 1 | |
| v8.4.190 | 1 | |
| 8.4.189 | 1 | |
| v8.4.187 | 0 | |
| v8.4.186 | 2 | |
| v8.4.184 | 2 | |
| 8.4.183 | 8 | |
| v8.4.176 | 8 | |
| 8.4.173 | 3 | |
| v8.4.171 |
Ultralytics 8.4.213
Large segmentation mask stacks now decode inside detection boxes, reducing memory and compute; tiled detections keep valid objects across class boundaries.
ultralytics 8.4.212
Ultralytics 8.4.212 adds OCR for Korean, Cyrillic, Arabic, Devanagari, Thai, Greek, Tamil, and Telugu scripts.
Ultralytics v8.4.211
Compiled SAM 3 image encoding is 7-9% faster, with INT8 ExecuTorch export and faster OCR on NVIDIA GPUs.
Ultralytics 8.4.210
OpenVINO and LiteRT now accept PyTorch-style device names: cpu, xpu, and npu.
Ultralytics 8.4.208
Prediction and tracking now default to 0.5 NMS IoU, while validation keeps 0.7.
Ultralytics 8.4.207
Ultralytics 8.4.207 fixes PaddlePaddle export on new installs and CUDA machines.
Ultralytics 8.4.206
Ultralytics 8.4.206 reads upside-down and bottom-to-top text in OCR using a text-line orientation classifier.
Ultralytics 8.4.204
High-resolution tiling is available across training, validation, and prediction, with FastSAM prompt, model export, and dataset handling improvements.
Ultralytics v8.4.203
Adds NXP i.MX95 Neutron NPU export for YOLO models via format=nxp.
Show the other 16 rows
Ultralytics v8.4.201
Ultralytics v8.4.201 makes PIL-based plotting up to 15x faster across all tasks.
Ultralytics 8.4.200
Arabic class names now display correctly in plots.
Ultralytics 8.4.199
Ultralytics 8.4.199 sends the raw run name to Ultralytics Platform.
Ultralytics 8.4.194
AutoBatch now sizes from the train dataset for every task, with fixes for model loading, results saving, datasets, and deployment workflows.
Ultralytics 8.4.193
CPU ONNX Runtime and OpenVINO prediction is up to 2x faster on some systems.
Ultralytics v8.4.192
v8.4.192 supports torch 2.15 CUDA in DDP training and ONNX Runtime inference.
Ultralytics 8.4.191
RKNN INT8 export now enabled for detection, segmentation, pose, and oriented bounding box models.
Ultralytics v8.4.190
YOLO26-seg training now uses Mask-IoU-aware label assignment, with fixes for export, inference, training utilities, and data handling.
Ultralytics 8.4.189
Ultralytics 8.4.189 fixes IMX500 and Hailo exports, dataset errors, and model output handling.
Ultralytics v8.4.187
YOLO26-pose training now uses OKS-only label assignment with a pyramid-level gate.
Ultralytics v8.4.186
Adds a shared import helper to load torch 1.10-1.12 and MNN on glibc 2.41+.
Ultralytics v8.4.184
v8.4.184 reads multi-page grayscale TIFFs as their first page unless every band is requested.
Ultralytics 8.4.183
Ultralytics 8.4.183 updates Axelera SDK to 1.9.0 and supports every YOLO26 task.
Ultralytics v8.4.176
YOLO-World checkpoints after set_classes drop from 328 MB to 26 MB by not caching the CLIP text encoder.
Ultralytics 8.4.173
Ultralytics 8.4.173 adds AMD Xilinx Vitis AI export for Versal AI Edge Gen 2 NPUs.
Ultralytics v8.4.171
Adds end-to-end AMD GPU support via ROCm and MIGraphX for training and native PyTorch mode.
One line every time AI actually ships, around the clock, free on Threads and Bluesky. Every one is in The Deploy Log with its source.