v2.2.1 1.11.0
roboflow/rf-detrSep 11, 2026 → Sep 24, 202614 days between releases
What's New
Adds OpenVINO, LiteRT, and Core AI export formats plus dynamic-batch TensorRT engines.
Breaking Changes
- Export internals restructured; update imports for direct converter functions (e.g., export_onnx -> OnnxExporter).
- TrainConfig.multi_scale is now a three-state enum; remove boolean do_random_resize_via_padding.
- TensorRT export with backbone_only=True now requires a -backbone suffix to prevent overwrites.
- Removed functions: RFDETR.optimize_for_inference(), TrainConfig.lr_drop, and TrainConfig.lr_min_factor.
New Features
- OpenVINO IR export with configurable precision (float32/float16).
- Apple Core AI export targeting iOS, iPadOS, and macOS 27+.
- Dynamic-batch TensorRT engines accepting a range of batch sizes.
- WebDataset streaming reader for tar-shard training input.
- Opt-in CUDA graph training replay for single-GPU detection.
Improvements
- Restructured export internals using a registry pattern.
- Training acceleration via CUDA graph replay and WebDataset streaming.
Upgrade Notes
Update import paths for direct converter functions (e.g., `export_onnx` -> `OnnxExporter`). Update `TrainConfig.multi_scale` to the new enum and remove the `do_random_resize_via_padding` parameter. Be aware that TensorRT engines with `backbone_only=True` now require a `-backbone` suffix.
The Two Releases
v2.2.1
Sep 11, 2026· 1 month ago
This release focuses on memory management enhancements with Auto Fin proactive reviews and the PowerContext backend, enhances the QwenPaw Creator with blueprint workbenches and Docker support, and refines the Console UI/UX alongside various security and compatibility fixes.
Full release notesv1.11.0: Wider export surface, faster training
Sep 24, 2026· 27 days ago
This release expands export support with OpenVINO IR, LiteRT, and Apple Core AI formats, introduces dynamic-batch TensorRT engines, and accelerates training via CUDA graph replay and WebDataset streaming.
Full release notes