Title: Ultralytics v8.3.179 — Stability Patch: Reliable Exports, Leaner Docker, YOLOE Usability Boost
Summary
A stability-focused release that pins TensorFlow for reliable exports, delivers export-optimized Docker images, improves YOLOE/World class handling, and tightens CI for fewer flaky runs. If you export to TensorFlow, please align your environment with tensorflow<=2.19.0 and numpy<2.0.0.
What’s New in v8.3.179
-
Export Reliability
- Pinned tensorflow<=2.19.0 and added numpy<2.0.0 in export extras to avoid known TF export issues.
- New Dockerfile-export providing preinstalled export dependencies and warm-started exports for faster, consistent testing.
-
CI and Workflows
- Temporarily disabled GPU CI to stabilize builds during server upgrades.
- Docker CI now tests with export-tagged images (latest-export, latest-python-export) to reflect real export environments.
- SBOM workflow installs package in editable mode for accurate dependency capture.
-
Docker Improvements
- Upgraded base image to pytorch/pytorch:2.8.0-cuda12.8-cudnn9-runtime.
- Leaner images with apt-get clean, trimmed caches, and split export deps into a dedicated image.
-
YOLOE/World Usability
- CLI now accepts classes=“person,bus” for YOLOE/World.
- set_classes auto-generates text embeddings if not provided, simplifying zero-shot/custom class usage.
-
Developer Quality-of-Life
- Removed pinned Click dependency; default compatible versions now used.
- Aligned yolo checks output for clearer logs.
New Features
- Export-ready Docker image and pipeline alignment for reproducible, fast exports across TF, ONNX, CoreML, OpenVINO, and TF.js.
Improvements
- More dependable exports through pinned dependency versions.
- Smaller, cleaner Docker images with clearer separation of concerns.
- Easier YOLOE/World class handling via CLI strings and auto-embeddings.
- Reduced CI flakiness and improved SBOM accuracy.
Bug Fixes
- Addressed TF export breakages by pinning tensorflow<=2.19.0 and numpy<2.0.0.
- Fixed spacing in yolo checks output for cleaner logs.
Purpose & Impact
- Dependable exports across multiple backends.
- Faster and more reproducible Docker-based workflows.
- Lower barrier to zero-shot and custom-class experiments with YOLOE/World.
- Reduced CI noise and costs, with GPU runs still available on demand.
- Better supply-chain transparency and simpler dev setup.
Pull Requests and Authors
- SBOM uv sync fix by @glenn-jocher — SBOM accuracy improvement
PR: SBOM uv sync fix - Use apt-get clean in Dockerfiles by @glenn-jocher — Leaner Docker images
PR: Use apt-get clean in Dockerfiles - YOLOE: Simplify set_classes and support text prompt CLI usage by @RizwanMunawar — Easier class handling
PR: YOLOE: Simplify set_classes and support text prompt CLI usage - chore: remove click version lock as 8.2.2 was yanked by @onuralpszr — Dependency stability
PR: remove click version lock as 8.2.2 was yanked - FROM pytorch/pytorch:2.8.0-cuda12.8-cudnn9-runtime by @glenn-jocher — Upgraded base image
PR: FROM pytorch/pytorch:2.8.0-cuda12.8-cudnn9-runtime - Disable GPU CI for server upgrades by @glenn-jocher — CI stabilization
PR: Disable GPU CI for server upgrades - Fix yolo checks spacing by @glenn-jocher — Log readability
PR: Fix yolo checks spacing - ultralytics 8.3.179 Pin tensorflow<=2.19.0 for exports by @glenn-jocher — Export reliability
PR: ultralytics 8.3.179 Pin tensorflow<=2.19.0 for exports
Links
- Release notes: Ultralytics v8.3.179
- Full changelog: Compare v8.3.178…v8.3.179
Getting Started
- Upgrade: pip install -U ultralytics
- Tip for TensorFlow exports: ensure tensorflow<=2.19.0 and numpy<2.0.0 to match this release.
- Try the new export-ready Docker images for faster, reproducible pipelines.
We’d love your feedback
Please try v8.3.179 and let us know how it works for your workflows—especially exports and YOLOE/World class prompts. Your input helps us keep YOLO fast, stable, and easy to use for everyone.
Thanks to our contributors and the YOLO community for helping make this release happen!