v1.9.0 1.11.0
0xNyk/council-of-high-intelligenceSep 22, 2026 → Sep 24, 20262 days between releases
What's New
New release adds OpenVINO, LiteRT, and Core AI export formats while restructuring export internals and removing deprecated training configuration options
Breaking Changes
- Export internals restructured into Exporter classes (direct imports like `export_onnx` are removed)
- TensorRT engine naming changed when using `backbone_only=True`
- `TrainConfig.multi_scale` is now a three-state enum and `do_random_resize_via_padding` is removed
- `RFDETR.optimize_for_inference()`, `TrainConfig.lr_drop`/`lr_min_factor`, and the standalone synthetic-dataset generator are removed
New Features
- OpenVINO IR export with configurable precision (float32/float16)
- Apple Core AI export for iOS/iPadOS/macOS 27+
- CUDA graph training replay for faster single-GPU detection
- WebDataset streaming reader for tar-shard training input
- Dynamic-batch TensorRT engines supporting a range of batch sizes
Improvements
- Restructured export internals into format-specific Exporter classes
- Dynamic-batch TensorRT engines built via a single optimization profile
- CUDA graph training replay combinable with `torch.compile`
- WebDataset streaming reader for scalable training input
Upgrade Notes
Code importing old converter functions directly needs updated imports. `TrainConfig.multi_scale` is now an enum, and `do_random_resize_via_padding` is removed. TensorRT engine naming changes when `backbone_only=True`. Consult the Migration guide for exact code examples.
The Two Releases
v1.9.0
Sep 22, 2026· 1 month ago
This release adds TikTok comment automation with private replies, WhatsApp in-app voice calling, and spintax support for flows. It also refactors the worker to move job handler data access into business services and improves the Threads UI visibility via an allowlist.
Full release notesv1.11.0: Wider export surface, faster training
Sep 24, 2026· 28 days ago
RF-DETR v1.11.0 widens the export surface by adding OpenVINO IR, LiteRT, and Apple Core AI formats while speeding up training with dynamic-batch TensorRT engines and CUDA graph replay.
Full release notes