# New Release: Ultralytics v8.4.105

**URL:** <https://community.ultralytics.com/t/new-release-ultralytics-v8-4-105/2135>\
**Category:** Discussion\
**Tags:** ultralytics-official, releases, announcements\
**Created:** [July 25, 2026, 1:32pm UTC](https://community.ultralytics.com/t/new-release-ultralytics-v8-4-105/2135 "2026-07-25T13:32:00Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**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:** [July 25, 2026, 1:32pm UTC](https://community.ultralytics.com/t/new-release-ultralytics-v8-4-105/2135/1 "2026-07-25T13:32:00Z")

</div>

# 🚀 Ultralytics v8.4.105 Released

Ultralytics **v8.4.105** is now available! This release makes distillation checkpoints more reliable, expands **YOLO26 depth deployment** to Hailo and Qualcomm QNN, improves CoreML and mobile exports, introduces opt-in CUDA `channels_last` support, and delivers substantial performance, training, Platform, and documentation updates.

Explore the complete [Ultralytics v8.4.105 release](https://github.com/ultralytics/ultralytics/releases/tag/v8.4.105) on GitHub.

## 🌟 Highlights

- 🛠 **Reliable distillation checkpoints:** Interrupted and mid-training checkpoints now support `export`, `predict`, and `val`.
- 🚀 **Broader YOLO26 depth deployment:** Hailo-8L and Qualcomm Snapdragon QNN devices are now supported.
- 🍎 **Better CoreML dynamic exports:** YOLO26 segmentation and depth models use export-compatible upsampling.
- 📱 **Consistent mobile input sizes:** Classification exports use `224`; detection, segmentation, semantic segmentation, depth, pose, and OBB use `640`.
- 🧠 **Optional CUDA `channels_last`:** Training and native PyTorch prediction can benefit from improved throughput while default NCHW behavior remains unchanged.
- ⚙ **Stronger fusion and quantization:** Fixes cover convolution layouts, ConvTranspose2d fusion, OpenVINO INT8 DFL operations, and CUDA calibration.
- 🌐 **Improved Ultralytics Platform workflows:** Depth API responses, configurable depth-map precision, annotation, activity export, billing, teams, deployment guidance, and API documentation have all been refined.

## ✨ New Features

- [Added Hailo export and inference for YOLO26 depth](https://github.com/ultralytics/ultralytics/pull/25348) by [@JESUSROYETH](https://github.com/JESUSROYETH).
- [Added gated CUDA `channels_last` support for training and prediction](https://github.com/ultralytics/ultralytics/pull/25362) by [@raimbekovm](https://github.com/raimbekovm).
- [Added cross-dataset class-row weight transfer for RT-DETR](https://github.com/ultralytics/ultralytics/pull/24423) by [@artest08](https://github.com/artest08).
- [Baked semantic argmax class maps into OpenVINO export graphs](https://github.com/ultralytics/ultralytics/pull/25395) by [@JESUSROYETH](https://github.com/JESUSROYETH).
- [Standardized mobile exports at classification `224` and other tasks at `640`](https://github.com/ultralytics/ultralytics/pull/25387) by [@glenn-jocher](https://github.com/glenn-jocher).

## ⚡ Performance and Workflow Improvements

- [Stride-aligned AutoBatch probe sizes for multi-scale training](https://github.com/ultralytics/ultralytics/pull/25352) by [@glenn-jocher](https://github.com/glenn-jocher).
- [Expanded depth-task fuzz coverage](https://github.com/ultralytics/ultralytics/pull/25357) by [@glenn-jocher](https://github.com/glenn-jocher).
- [Kept DFL operations in floating point during OpenVINO INT8 quantization](https://github.com/ultralytics/ultralytics/pull/25306) by [@synml](https://github.com/synml).
- [Reduced `Heatmap` accumulator memory to a single channel](https://github.com/ultralytics/ultralytics/pull/25342) by [@JESUSROYETH](https://github.com/JESUSROYETH).
- [Removed an unnecessary validator device round trip](https://github.com/ultralytics/ultralytics/pull/25344) by [@JESUSROYETH](https://github.com/JESUSROYETH).
- [Removed redundant feature-map cloning from `WorldDetect.forward`](https://github.com/ultralytics/ultralytics/pull/25345) by [@JESUSROYETH](https://github.com/JESUSROYETH).
- [Vectorized tracker `multi_predict` and `multi_gmc` loops](https://github.com/ultralytics/ultralytics/pull/25343) by [@JESUSROYETH](https://github.com/JESUSROYETH).
- [Consolidated dataset label caching with shared templates and per-class hooks](https://github.com/ultralytics/ultralytics/pull/25359) by [@Laughing-q](https://github.com/Laughing-q).
- [Optimized `BboxLoss` weights and pose strides for matched anchors](https://github.com/ultralytics/ultralytics/pull/25346) by [@JESUSROYETH](https://github.com/JESUSROYETH).
- [Simplified convolution and BatchNorm fusion broadcasting](https://github.com/ultralytics/ultralytics/pull/25367) by [@glenn-jocher](https://github.com/glenn-jocher).
- [Gathered end-to-end top-k results with zero-copy expansion](https://github.com/ultralytics/ultralytics/pull/25369) by [@JESUSROYETH](https://github.com/JESUSROYETH).
- [Built TAL box-metric gather indices directly on the target device](https://github.com/ultralytics/ultralytics/pull/25397) by [@raimbekovm](https://github.com/raimbekovm).
- [Vectorized GMC point filtering with boolean masks](https://github.com/ultralytics/ultralytics/pull/25398) by [@raimbekovm](https://github.com/raimbekovm).
- [Skipped redundant float32 input copies in OpenVINO and Paddle backends](https://github.com/ultralytics/ultralytics/pull/25413) by [@JESUSROYETH](https://github.com/JESUSROYETH).

## 🛠 Bug Fixes and Reliability

- [Fixed CoreML dynamic segmentation upsampling](https://github.com/ultralytics/ultralytics/pull/25347) by [@glenn-jocher](https://github.com/glenn-jocher).
- [Fixed CoreML dynamic depth bilinear upsampling](https://github.com/ultralytics/ultralytics/pull/25350) by [@glenn-jocher](https://github.com/glenn-jocher).
- [Aligned ONNX-backed export arguments across formats](https://github.com/ultralytics/ultralytics/pull/25356) by [@amanharshx](https://github.com/amanharshx).
- [Fixed convolution and BatchNorm fusion for `channels_last` weights](https://github.com/ultralytics/ultralytics/pull/25363) by [@raimbekovm](https://github.com/raimbekovm).
- [Fixed the ConvTranspose2d and BatchNorm fusion weight axis](https://github.com/ultralytics/ultralytics/pull/25364) by [@raimbekovm](https://github.com/raimbekovm).
- [Fixed CUDA calibration crashes caused by inference tensors](https://github.com/ultralytics/ultralytics/pull/25351) by [@JESUSROYETH](https://github.com/JESUSROYETH).
- [Fixed single-class depth training](https://github.com/ultralytics/ultralytics/pull/25370) by [@glenn-jocher](https://github.com/glenn-jocher).
- [Fixed RMSprop incorrectly constructing MuSGD](https://github.com/ultralytics/ultralytics/pull/25376) by [@raimbekovm](https://github.com/raimbekovm).
- [Restricted MuSGD resolution to the explicit optimizer path](https://github.com/ultralytics/ultralytics/pull/25383) by [@raimbekovm](https://github.com/raimbekovm).
- [Updated compatibility for Ruff 0.16.0](https://github.com/ultralytics/ultralytics/pull/25390) by [@glenn-jocher](https://github.com/glenn-jocher).
- [Corrected SAM3 semantic-mask thresholding](https://github.com/ultralytics/ultralytics/pull/25386) by [@Y-T-G](https://github.com/Y-T-G).
- [Preserved falsey activation arguments](https://github.com/ultralytics/ultralytics/pull/25391) by [@glenn-jocher](https://github.com/glenn-jocher).
- [Fixed semantic targets for empty YOLO26 segmentation images](https://github.com/ultralytics/ultralytics/pull/25392) by [@JESUSROYETH](https://github.com/JESUSROYETH).
- [Recorded effective CoreML NMS precision](https://github.com/ultralytics/ultralytics/pull/25389) by [@amanharshx](https://github.com/amanharshx).
- [Improved GMC outlier filtering on both sides of the mean](https://github.com/ultralytics/ultralytics/pull/25404) by [@raimbekovm](https://github.com/raimbekovm).
- [Padded every multispectral channel during `RandomPerspective` warps](https://github.com/ultralytics/ultralytics/pull/25411) by [@JESUSROYETH](https://github.com/JESUSROYETH).
- [Fixed `TRACKTRACK.update` for a single torch-backed detection](https://github.com/ultralytics/ultralytics/pull/25394) by [@JESUSROYETH](https://github.com/JESUSROYETH).
- [Fixed the `intel:gpu` alias on multi-GPU OpenVINO systems](https://github.com/ultralytics/ultralytics/pull/25393) by [@JESUSROYETH](https://github.com/JESUSROYETH).
- [Derived export argument validation from the format table](https://github.com/ultralytics/ultralytics/pull/25368) by [@amanharshx](https://github.com/amanharshx).
- [Validated pose, `kobj`, RLE, angle, and erasing options at the configuration layer](https://github.com/ultralytics/ultralytics/pull/25384) by [@usegitdivyansh](https://github.com/usegitdivyansh).
- [Fixed NCNN batched inference dropping images after the first](https://github.com/ultralytics/ultralytics/pull/25412) by [@JESUSROYETH](https://github.com/JESUSROYETH).
- [Added `fuse()` delegation for distillation checkpoints](https://github.com/ultralytics/ultralytics/pull/25415) by [@glenn-jocher](https://github.com/glenn-jocher).

## 🌐 Platform, Export, and Documentation

- [Documented YOLO26 depth in the Platform documentation and API reference](https://github.com/ultralytics/ultralytics/pull/25349) by [@glenn-jocher](https://github.com/glenn-jocher).
- [Documented depth-map bit precision and model-resolution maps](https://github.com/ultralytics/ultralytics/pull/25354) by [@glenn-jocher](https://github.com/glenn-jocher).
- [Added Platform guidance following export failures](https://github.com/ultralytics/ultralytics/pull/25353) by [@glenn-jocher](https://github.com/glenn-jocher).
- [Added a depth workflow video to the documentation](https://github.com/ultralytics/ultralytics/pull/25374) by [@RizwanMunawar](https://github.com/RizwanMunawar).
- [Clarified the YOLOE export workflow for C++ ONNX Runtime inference](https://github.com/ultralytics/ultralytics/pull/25373) by [@23z156-wq](https://github.com/23z156-wq).
- [Added CPU ONNX and T4 TensorRT 10 speed columns to depth benchmarks](https://github.com/ultralytics/ultralytics/pull/25382) by [@Bovey0809](https://github.com/Bovey0809).
- [Added stable Platform screenshot markers](https://github.com/ultralytics/ultralytics/pull/25388) by [@glenn-jocher](https://github.com/glenn-jocher).
- [Added a T4 TensorRT comparison with Depth Anything V2](https://github.com/ultralytics/ultralytics/pull/25399) by [@raimbekovm](https://github.com/raimbekovm).
- [Fixed shared export-table rendering](https://github.com/ultralytics/ultralytics/pull/25401) by [@glenn-jocher](https://github.com/glenn-jocher).
- [Fixed GMC documentation examples](https://github.com/ultralytics/ultralytics/pull/25402) by [@raimbekovm](https://github.com/raimbekovm).
- [Documented class-alias weight transfer for fine-tuning](https://github.com/ultralytics/ultralytics/pull/25409) by [@artest08](https://github.com/artest08).
- [Added depth-map colorization options to the documentation](https://github.com/ultralytics/ultralytics/pull/25400) by [@onuralpszr](https://github.com/onuralpszr).
- [Updated reCamera links and contribution principles](https://github.com/ultralytics/ultralytics/pull/25380) by [@onuralpszr](https://github.com/onuralpszr).

## 🧰 CI and Repository Maintenance

- [Removed duplicate Python setup from CI](https://github.com/ultralytics/ultralytics/pull/25366) by [@glenn-jocher](https://github.com/glenn-jocher).
- [Built the Docker runner before the export image](https://github.com/ultralytics/ultralytics/pull/25371) by [@glenn-jocher](https://github.com/glenn-jocher).
- [Updated the README banner for YOLO Vision 2026 registration](https://github.com/ultralytics/ultralytics/pull/25377) by [@raimbekovm](https://github.com/raimbekovm).
- [Updated remaining banner links for YOLO Vision 2026 registration](https://github.com/ultralytics/ultralytics/pull/25381) by [@raimbekovm](https://github.com/raimbekovm).
- [Simplified the `AGENTS.md` core principles](https://github.com/ultralytics/ultralytics/pull/25414) by [@glenn-jocher](https://github.com/glenn-jocher).

## 🙌 New Contributors

A warm welcome to our first-time contributors:

- [@23z156-wq](https://github.com/23z156-wq), who contributed the [YOLOE C++ ONNX Runtime export documentation](https://github.com/ultralytics/ultralytics/pull/25373).
- [@usegitdivyansh](https://github.com/usegitdivyansh), who contributed [configuration-layer argument validation](https://github.com/ultralytics/ultralytics/pull/25384).

Thank you to every contributor and the wider YOLO community for helping improve this release! 💙

## 📦 Upgrade Today

Install or upgrade to v8.4.105 with:

```bash
pip install -U ultralytics

```

For the complete commit-level comparison, review the [v8.4.104 to v8.4.105 changelog](https://github.com/ultralytics/ultralytics/compare/v8.4.104...v8.4.105).

Give the release a try and share your feedback, benchmark results, and deployment experiences with the community! 🚀