New Release: Ultralytics v8.3.241

Discussion
, ,
1

:rocket: Ultralytics v8.3.241 is live (macOS speedups + export stability) :green_apple::package:

Quick overview: v8.3.241 boosts ONNX Runtime performance on macOS by preferring CoreML on Apple hardware, improves ExecuTorch export reliability, adds Rockchip RV1126B support, and polishes docs/tooling. :high_voltage:

You can review the release details in the Ultralytics v8.3.241 release page and the full changelog comparison.

:glowing_star: New Features / Highlights

:green_apple: macOS ONNX Runtime: Prefer CoreML on Apple hardware

Ultralytics now prefers CoreMLExecutionProvider (when available) on macOS when using device="mps", rather than falling back to CPU.

  • Provider selection is now CUDA > CoreML (MPS) > CPU, with clearer logging
  • IO binding is only used where supported (CUDA), avoiding CoreML issues

PR: ORT CoreML execution provider when device="mps" by @glenn-jocher

:package: Export & Backend Improvements

:puzzle_piece: ExecuTorch export reliability improvements

  • export_executorch() now pins numpy<=2.3.5 to avoid known coremltools failures
  • Removed a redundant import torch in the exporter path

PRs:

:wrench: TF/SavedModel & TensorRT backend cleanups

  • TF SavedModel loading simplified to always use tf.saved_model.load in this path
  • TensorRT engine parsing refactored for clearer handling across TRT <10 and TRT ≥10

PR: Fix macros by @glenn-jocher

:brick: Edge Device Support

:white_check_mark: Rockchip RV1126B support

  • Added "rv1126b" to supported Rockchip chips
  • Docs updated to explicitly mention RV1126B

PRs:

:framed_picture: Visualization & Data Handling

:rainbow: 2‑channel image support in Annotator

ultralytics.utils.plotting.Annotator now handles 2‑channel numpy images by converting to 3‑channel before drawing/saving.

PR: Add support for handling 2-channel images in Annotator class by @kenanking

:books: Docs & Tooling Polish

:test_tube: Try it now

Upgrade:

pip install -U ultralytics

Quick test (example with Ultralytics YOLO + ONNX on macOS):

from ultralytics import YOLO

model = YOLO("yolo11n.pt")
model.export(format="onnx")
model.predict("image.jpg", device="mps")

:raising_hands: New Contributors

Welcome, and thank you for strengthening the YOLO community! :tada:

:speech_balloon: Feedback welcome

Please try v8.3.241 and share results—especially:

  • macOS ONNX Runtime + device="mps" performance :green_apple::high_voltage:
  • ExecuTorch export stability :package:
  • Rockchip RV1126B deployments :brick:

If anything looks off, opening an issue with logs + minimal repro steps helps us fix it fast.