# New Release: Ultralytics v8.4.128 **URL:** \ **Category:** Discussion\ **Tags:** releases, announcements, ultralytics-official\ **Created:** [August 25, 2026, 1:20pm UTC](https://community.ultralytics.com/t/new-release-ultralytics-v8-4-128/2185 "2026-08-25T13:20:11Z")\ **Posts on this page:** 1\ **Page:** 1
**Author:** ![glenn-jocher](https://sea1.discourse-cdn.com/flex001/user_avatar/community.ultralytics.com/glenn-jocher/32/83_2.png) [@glenn-jocher](https://community.ultralytics.com/u/glenn-jocher)\ **Post date:** [August 25, 2026, 1:20pm UTC](https://community.ultralytics.com/t/new-release-ultralytics-v8-4-128/2185/1 "2026-08-25T13:20:11Z")
# 🚀 Ultralytics v8.4.128 Released Ultralytics **v8.4.128** is here with more reliable OpenVINO batch inference, optimized TensorRT and ONNX exports, lower training RAM usage, faster SAM mask generation, safer FastSAM prompts, and stronger dataset and training utilities. 🌟 The [v8.4.128 GitHub release](https://github.com/ultralytics/ultralytics/releases/tag/v8.4.128) includes the complete release package and upgrade details. ## ⚡ Deployment and Export Improvements - In [PR #25921](https://github.com/ultralytics/ultralytics/pull/25921), [@glenn-jocher](https://github.com/glenn-jocher) introduced synchronous OpenVINO batch inference with a consistent `LATENCY` performance hint, reducing the risk of hangs from asynchronous dynamic INT8 workloads. - In [PR #25920](https://github.com/ultralytics/ultralytics/pull/25920), [@onuralpszr](https://github.com/onuralpszr) scoped TensorRT INT8 Sigmoid precision exclusions to detection heads for TensorRT 7–10 and removed the unnecessary exclusion for TensorRT 11. - In [PR #25915](https://github.com/ultralytics/ultralytics/pull/25915), [@Y-T-G](https://github.com/Y-T-G) capped ONNX exports at opset 18 to improve ONNX Runtime CUDA kernel coverage and avoid unwanted CPU fallback. - In [PR #24703](https://github.com/ultralytics/ultralytics/pull/24703), [@lakshanthad](https://github.com/lakshanthad) improved RKNN INT8 validation, fixed “no detections” behavior, and updated YOLO26 Rockchip benchmarks. - In [PR #23187](https://github.com/ultralytics/ultralytics/pull/23187), [@acgist](https://github.com/acgist) added class-aware NMS to shared C++ postprocessing. > **Compatibility note:** OpenVINO now prioritizes predictable, latency-safe synchronous execution. Throughput-focused users should benchmark v8.4.128 on their specific hardware and workloads. ## 🧠 Performance and Memory - In [PR #24673](https://github.com/ultralytics/ultralytics/pull/24673), [@raimbekovm](https://github.com/raimbekovm) moved `cache='ram'` images into a shared contiguous buffer, preventing DataLoader workers from duplicating the cache during forked training. - In [PR #25909](https://github.com/ultralytics/ultralytics/pull/25909), [@JESUSROYETH](https://github.com/JESUSROYETH) optimized SAM auto-mask generation by encoding each crop once and reusing its features across point batches. - In [PR #23664](https://github.com/ultralytics/ultralytics/pull/23664), [@gabe-zhang](https://github.com/gabe-zhang) moved YOLO26n AMP-check weights into the global user configuration directory so they can be reused across projects. ## 🛠 Training and Data Fixes - In [PR #22724](https://github.com/ultralytics/ultralytics/pull/22724), [@Y-T-G](https://github.com/Y-T-G) prevented externally defined `RANK` and `LOCAL_RANK` variables from incorrectly activating DDP behavior. - In [PR #25922](https://github.com/ultralytics/ultralytics/pull/25922), [@glenn-jocher](https://github.com/glenn-jocher) fixed Ray Tune metric aggregation across multiple datasets and corrected completed-epoch reporting. - In [PR #25906](https://github.com/ultralytics/ultralytics/pull/25906), [@aswanth-07](https://github.com/aswanth-07) preserved bounding-box formats during indexing. - In [PR #25907](https://github.com/ultralytics/ultralytics/pull/25907), [@aswanth-07](https://github.com/aswanth-07) preserved instance axes during scalar indexing. - In [PR #25912](https://github.com/ultralytics/ultralytics/pull/25912), [@ShuaiLYU](https://github.com/ShuaiLYU) made ground-truth side clamping monotonic for more consistent candidate selection with very small boxes. - In [PR #25916](https://github.com/ultralytics/ultralytics/pull/25916), [@JESUSROYETH](https://github.com/JESUSROYETH) clipped FastSAM box and point prompts to image boundaries, preventing invalid coordinates from selecting incorrect masks. - In [PR #24357](https://github.com/ultralytics/ultralytics/pull/24357), [@Y-T-G](https://github.com/Y-T-G) added support for local image paths during NDJSON dataset conversion. - In [PR #20870](https://github.com/ultralytics/ultralytics/pull/20870), [@roel-klein](https://github.com/roel-klein) removed an unreachable `YOLODataset` task assertion. - In [PR #24163](https://github.com/ultralytics/ultralytics/pull/24163), [@noorchauhan](https://github.com/noorchauhan) removed an unreachable resume-training branch. ## 📚 Documentation and Usability - In [PR #18658](https://github.com/ultralytics/ultralytics/pull/18658), [@ambitious-octopus](https://github.com/ambitious-octopus) refreshed the Docker quickstart and added clearer local build instructions. - In [PR #24527](https://github.com/ultralytics/ultralytics/pull/24527), [@AffanBinFaisal](https://github.com/AffanBinFaisal) added task-specific K-Fold split guidance. - In [PR #25910](https://github.com/ultralytics/ultralytics/pull/25910), [@RizwanMunawar](https://github.com/RizwanMunawar) added a new instructional video to the documentation. - In [PR #23578](https://github.com/ultralytics/ultralytics/pull/23578), [@PhuTd03](https://github.com/PhuTd03) corrected the `InfiniteDataLoader` iteration documentation. - In [PR #22849](https://github.com/ultralytics/ultralytics/pull/22849), [@blm25](https://github.com/blm25) documented Android ExecuTorch runtime setup. - In [PR #25919](https://github.com/ultralytics/ultralytics/pull/25919), [@raimbekovm](https://github.com/raimbekovm) corrected augmentation parameters and clarified COCO JSON training, caching, and validation guidance. - In [PR #25908](https://github.com/ultralytics/ultralytics/pull/25908), [@raimbekovm](https://github.com/raimbekovm) clarified that tracking is a **mode** used with detection, segmentation, pose, or OBB models rather than a separate task. ## 👋 Welcome, New Contributors A warm welcome to first-time contributors [@roel-klein](https://github.com/roel-klein), [@acgist](https://github.com/acgist), [@blm25](https://github.com/blm25), [@noorchauhan](https://github.com/noorchauhan), [@PhuTd03](https://github.com/PhuTd03), [@aswanth-07](https://github.com/aswanth-07), and [@gabe-zhang](https://github.com/gabe-zhang). Thank you for helping improve Ultralytics YOLO! 🙌 ## 📦 Try v8.4.128 Upgrade to the latest release with: ```bash pip install -U ultralytics ``` Please test your training, export, and deployment workflows—especially OpenVINO batch inference, TensorRT INT8, ONNX Runtime CUDA, and RAM-cached training—and share your feedback with the community. For every change since v8.4.127, review the [full v8.4.128 changelog](https://github.com/ultralytics/ultralytics/compare/v8.4.127...v8.4.128) before upgrading production environments. 🚀