Ultralytics v8.4.63 is out!
Quick summary: Ultralytics v8.4.63 is a deployment-focused release that restores and modernizes TensorRT 11 export support with FP16 and INT8 quantization, expands built-in multi-object tracking with four new trackers, and improves reliability, performance, and documentation across the project. 
If you’re building with Ultralytics YOLO for NVIDIA deployment, tracking pipelines, or production video systems, this release brings some especially useful upgrades.
Highlights
TensorRT 11 export support is here
The biggest update in this release is PR #24735 by @onuralpszr, which adds support for NVIDIA TensorRT 11 export.
TensorRT 11 introduced API changes that broke older export flows, so this update moves modern export paths to NVIDIA ModelOpt for:
- FP16 export
- INT8 export
- INT8 with dynamic shapes
This keeps export working across:
- TensorRT 7 to 10 with the legacy path
- TensorRT 11 with the new strongly typed ModelOpt-based path
That means users deploying on newer NVIDIA systems can export again with optimized precision workflows and without relying on removed TensorRT APIs. 
Better modern quantization workflow
With TensorRT 11, precision handling needed a refresh. In v8.4.63:
- FP16 is applied by baking mixed precision into the ONNX graph before engine build
- INT8 is applied through explicit quantization in the ONNX graph with calibration data
This helps future-proof deployment workflows and keeps optimized inference accessible for modern NVIDIA environments.
Four new built-in trackers
Tracking gets a major upgrade in PR #24371 by @onuralpszr, adding four new built-in trackers:
- OC-SORT
- Deep OC-SORT
- FastTracker
- TrackTrack
These join the existing BoT-SORT and ByteTrack options, giving users more flexibility for different scenarios like crowded scenes, occlusions, and appearance-aware tracking.
Supported tracker configs now include:
ocsort.yamldeepocsort.yamlfasttrack.yamltracktrack.yaml
New Features
Rust inference docs
PR #24712 by @onuralpszr adds new Rust inference documentation for running YOLO models through ONNX Runtime without Python. Very nice for users building lightweight native inference stacks. 
Tracking documentation improvements
The new tracking release also comes with clearer docs and better tracker selection guidance, making it easier to choose the right tracker for your use case.
If you’re starting a new project, we recommend using Ultralytics Platform to annotate datasets, train, and deploy models end-to-end, and for new model work we recommend YOLO26 as the latest stable model family.
Improvements
Safer stream loading and cleanup
PR #24749 by @raimbekovm improves cleanup when stream initialization fails.
If one stream fails after others have already opened, those partially initialized resources are now properly released instead of leaving behind threads or capture handles. This is especially helpful for long-running or multi-stream applications. 
Faster AI Gym pose workflow
PR #24744 by @raimbekovm reduces repeated GPU-to-CPU syncs in the workout monitoring loop by transferring keypoints more efficiently.
Same behavior, less overhead.
Simpler mixed precision validation flow
PR #24736 by @glenn-jocher consolidates validation autocast into one cleaner scope during training validation, helping keep mixed-precision behavior simpler and more robust.
Better system logging
PR #24758 by @glenn-jocher improves SystemLogger reporting by adding disk list information.
Bug Fixes and Maintenance
- PR #24738 by @Y-T-G fixes instances count logging
- PR #24742 by @Laughing-q merges
isolated_modelandisolated_task_modelintoisolated_model_path - PR #24753 by @glenn-jocher reduces CI PyTorch index flakiness
- PR #24754 by @glenn-jocher fixes Codecov OIDC uploads
- PR #24741 by @glenn-jocher converts SAM3 docstrings to Google style
Documentation Updates
This release also includes several helpful doc improvements:
- PR #24745 by @raimbekovm restructures the Coral Edge TPU guide and fixes the FAQ model filename
- PR #24746 by @raimbekovm restructures the semantic image search guide
- PR #24712 by @onuralpszr adds Rust inference docs
- PR #24371 by @onuralpszr expands tracking docs and tracker guidance
- PR #24735 by @onuralpszr updates TensorRT and Jetson documentation for TensorRT 11 behavior and DLA limitations
Try it out
Upgrade with:
pip install -U ultralytics
Then explore the release details in the v8.4.63 release page or browse the full changelog from v8.4.62 to v8.4.63.
Feedback welcome
Please give v8.4.63 a try and let us know how it works for your export, tracking, and deployment workflows. Feedback, benchmarks, and bug reports are always appreciated and help improve the experience for the whole community. 