Ultralytics v8.3.239 is live!
Quick overview: v8.3.239 makes SAM 3 text prompting truly plug-and-play
, adds real-time disk/network throughput to SystemLogger()
, cleans up noisy warnings
, and broadens ONNX export support (including Jetson JetPack 6)
—plus improved contributor attribution in the docs
.
Upgrade anytime with pip install -U ultralytics.
New Features
SAM 3 text prompting is now plug‑and‑play 
Text prompting with SAM 3 is simpler and more standard:
- Switched from the custom SAM3 tokenizer to CLIP’s built-in
SimpleTokenizer() - Removed the need to download/pass
bpe_simple_vocab_16e6.txt.gz SAM3SemanticPredictorandSAM3VideoSemanticPredictorno longer acceptbpe_path- If
clipisn’t installed, Ultralytics will auto-install it from the Ultralytics CLIP repo - Docs and navigation updated accordingly
Implemented in PR Use CLIP tokenizer functionality and remove the need of bpe_simple_vocab for SAM3 by @Y-T-G.
Minimal example:
from ultralytics import SAM
model = SAM("sam3.pt")
results = model.predict("image.jpg", texts=["person", "dog"])
Improvements
SystemLogger() now supports real-time I/O throughput
SystemLogger.get_metrics() now supports rates=True to report disk/network MB/s:
- Disk:
read_mbs,write_mbs - Network:
recv_mbs,sent_mbs
This is especially helpful for spotting dataset loading bottlenecks or slow storage/networked filesystems.
Added in PR SystemLogger rate mode by @glenn-jocher, plus follow-up updates in PR SAM3 and SystemLogger() improvements by @glenn-jocher.
from ultralytics.utils import SystemLogger
logger = SystemLogger()
metrics = logger.get_metrics(rates=True)
print(metrics)
Cleaner, more focused logs 
Warning suppression is now centralized to reduce repeated noise during training/export runs, and TensorBoard graph logging is more resilient with a clearer fallback + single informative warning.
Shipped in PR Centralize warning suppression by @glenn-jocher.
ONNX export support expanded (including Jetson JetPack 6) 
ONNX requirements were updated to support newer versions (up to <2.0.0, including 1.20.0), removing older caps and simplifying Jetson environments.
Delivered in PR feat:
update onnx version to 1.20.0 by @onuralpszr.
Bug Fixes / Maintenance
Docs & contributor attribution improvements 

Docs now better credit contributors with expanded GitHub author mappings, avatars, aliases, and team mappings.
Merged in PR Restore GitHub authors with avatars and usernames by @glenn-jocher.
CI upkeep 
GitHub Actions updated: actions/download-artifact bumped v6 → v7.
Handled in PR Bump actions/download-artifact from 6 to 7 by @dependabot[bot].
Release links
You can review the full release details in the Ultralytics v8.3.239 release page, and browse all changes in the v8.3.238…v8.3.239 full changelog.
Try it & share feedback
Give v8.3.239 a spin—especially if you use SAM 3 prompting, TensorBoard, or ONNX export on Jetson—and let us know what you find (wins, regressions, or ideas). Community feedback helps keep Ultralytics YOLO sharp for everyone.