# New Release: Ultralytics v8.4.121

**URL:** <https://community.ultralytics.com/t/new-release-ultralytics-v8-4-121/2175>\
**Category:** Discussion\
**Tags:** ultralytics-official, releases, announcements\
**Created:** [August 18, 2026, 1:18pm UTC](https://community.ultralytics.com/t/new-release-ultralytics-v8-4-121/2175 "2026-08-18T13:18:58Z")\
**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:** [August 18, 2026, 1:18pm UTC](https://community.ultralytics.com/t/new-release-ultralytics-v8-4-121/2175/1 "2026-08-18T13:18:58Z")

</div>

# 🚀 Ultralytics v8.4.121 Released

Ultralytics **v8.4.121** is now available! This release improves **OpenVINO INT8 export reliability for YOLO26** , significantly expands **Ultralytics Platform API and workflow documentation** , and strengthens security, CI, and dependency support.

Explore the [v8.4.121 GitHub release](https://github.com/ultralytics/ultralytics/releases/tag/v8.4.121) for the complete release details.

## 🌟 Release Highlights

### ⚡ More Reliable OpenVINO INT8 Exports

The most important update in v8.4.121 fixes OpenVINO INT8 detection-head handling for YOLO26 models.

Instead of relying on fragile PyTorch layer matching, quantization now uses exact operation names from the converted OpenVINO graph. This ensures that Detect decoding, DFL, and Sigmoid operations remain in floating point while preserving strict NNCF validation.

Testing confirmed a successful **YOLO26n-P2 INT8 export with 55 exact ignored operations** and no unwanted `FakeQuantize` nodes.

This fix was contributed by [@glenn-jocher](https://github.com/glenn-jocher) in [PR #25841](https://github.com/ultralytics/ultralytics/pull/25841), directly addressing reported export failures and improving quantization predictability.

## 🏷 Platform API, Dataset, and Annotation Improvements

The Ultralytics Platform documentation now more closely reflects the current product and API contract, making it easier to automate complete computer vision workflows.

Key updates include:

- Generated `ultralytics-platform` Python SDK documentation alongside REST examples
- Updated endpoints, authentication, pagination, rate limits, response formats, and OpenAPI guidance
- Expanded coverage for datasets, images, exports, training, deployments, storage, billing, usage, trash, and accounts
- Clearer URL, cloud-storage, and On Premise dataset import guidance
- COCO and NDJSON import documentation
- Class merging, deletion, conflict handling, clustering, readiness checks, and version restore guidance
- Expanded annotation visibility, copy-and-paste workflows, and keyboard shortcuts
- Clarification that Pascal VOC XML labels are not imported
- Clearer workspace API-key permissions and ownership requirements

The broader Platform documentation refresh was delivered by [@glenn-jocher](https://github.com/glenn-jocher) in [PR #25808](https://github.com/ultralytics/ultralytics/pull/25808), while generated SDK coverage was added in [PR #25823](https://github.com/ultralytics/ultralytics/pull/25823) by the same contributor.

## 💳 Account, Billing, and Team Documentation

This release also improves explanations for:

- Credit metering, monthly expiration, and automatic top-ups
- Seat billing, renewals, and downgrade effects
- Workspace roles and owner-only API keys
- Invitations, seat reuse, ownership transfer, and team deletion
- Activity exports, trash permissions, storage usage, and account deletion

A Platform account-settings spelling correction was contributed by [@UltralyticsAssistant](https://github.com/UltralyticsAssistant) in [PR #25821](https://github.com/ultralytics/ultralytics/pull/25821), improving documentation quality.

## 🌐 Deployment and Inference Documentation

Deployment documentation now provides clearer guidance for:

- Dedicated endpoint lifecycle operations
- Authentication and rate limits
- Model replacement and capacity behavior
- Health checks, metrics, and logs
- Video inference and depth response options
- Endpoint-specific API references
- Generated deployment URLs

These updates make it easier to move from dataset preparation and training to production inference using the Ultralytics Platform.

## 🔒 Security and CI Improvements

Several changes strengthen build security and CI reliability:

- Checkout credentials are no longer copied into Docker images, implemented by [@glenn-jocher](https://github.com/glenn-jocher) in [PR #25805](https://github.com/ultralytics/ultralytics/pull/25805).
- GitHub configuration was standardized by [@glenn-jocher](https://github.com/glenn-jocher) in [PR #25822](https://github.com/ultralytics/ultralytics/pull/25822).
- SlowTests attempts were reduced from 180 to 120 minutes while retaining one retry, contributed by [@glenn-jocher](https://github.com/glenn-jocher) in [PR #25839](https://github.com/ultralytics/ultralytics/pull/25839).
- Self-hosted runner cleanup was updated to v1.4.38 by [@UltralyticsAssistant](https://github.com/UltralyticsAssistant) in [PR #25811](https://github.com/ultralytics/ultralytics/pull/25811), then to v1.4.39 in [PR #25836](https://github.com/ultralytics/ultralytics/pull/25836) by the same contributor.
- Dependabot GitHub Actions checks now run weekly instead of daily.

Together, these changes reduce credential exposure risks, stale runner state, and excessively long CI jobs. 🛡

## 📚 Dependencies and Documentation Maintenance

Additional updates include:

- The `setuptools` upper bound was increased from `<81.0.0` to `<85.0.0` by [@dependabot[bot]](https://github.com/dependabot) in [PR #25826](https://github.com/ultralytics/ultralytics/pull/25826).
- NNCF support was expanded from `<3.0.0,>=2.14.0` to `>=2.14.0,<4.0.0` by [@dependabot[bot]](https://github.com/dependabot) in [PR #25825](https://github.com/ultralytics/ultralytics/pull/25825).
- Rust inference examples now use `ultralytics-inference` 0.0.35, updated by [@onuralpszr](https://github.com/onuralpszr) in [PR #25818](https://github.com/ultralytics/ultralytics/pull/25818).
- Supported-task banners and task enumerations were corrected to represent YOLO26 coverage consistently, contributed by [@raimbekovm](https://github.com/raimbekovm) in [PR #25807](https://github.com/ultralytics/ultralytics/pull/25807).
- Trainer changes from [@AyushExel](https://github.com/AyushExel) are included through [PR #24](https://github.com/ultralytics/ultralytics/pull/24).

## 📦 Upgrade Today

Install or upgrade to the latest release with:

```bash
pip install -U ultralytics

```

YOLO26 is the latest stable and recommended Ultralytics YOLO generation for new projects, with all model sizes and detect, segment, classify, pose, and OBB tasks supported.

Review the [full v8.4.120 to v8.4.121 changelog](https://github.com/ultralytics/ultralytics/compare/v8.4.120...v8.4.121) to inspect every change.

Please try v8.4.121—especially OpenVINO INT8 export with YOLO26—and share your results, questions, or regressions with the Ultralytics community. Thanks to every contributor who helped make this release possible! 🙌