Cut AdamW step time by 85% and a full Nano training step by 23.5% by grouping parameters by learning rate and weight decay (465 parameter groups down to 28), with bit-identical weights. #1409

Rebuilt the Hungarian matcher's compact path, padding to the largest per-image target count and copying only diagonal cost blocks to the host, cutting matcher time by 51–77% and peak CUDA memory by up to 76%. #1281 · #1297

Made compile=True actually compile: it crashed with explicit single-scale training and was silently a no-op under the default multi-scale setting. #1411 · #1436

Brought RF-DETR training to Cloud TPU (experimental), fixing multi-device XLA strategy selection, padding targets to a fixed shape to avoid per-batch recompiles, and making the XLA test suite pass on real TPU silicon. #1426 · #1427 · #1450

Reduced the DataLoader handoff from 114 objects to 9 per batch, improving throughput by up to 1.67× and preventing file-descriptor crashes at high worker counts. #1399

Added an opt-in Inductor inference backend and removed all seven decoder graph breaks, cutting batch-1 predict() latency by 47.7% on an L4. #1425

Corrected the encoder mask-loss computation across all released segmentation models and added the missing tests for SegmentationHead. #1331

Extended Ultralytics' Hailo backend from detection-only to seven tasks across the complete .pt → ONNX → INT8 → HEF export path. #25254 · #25259 · #25276 · #25280 · #25283 · #25348

Fixed RKNN INT8 exports across detection, segmentation, pose, and OBB that produced valid files but no detections. #25524

Made channels-last the automatic default for native PyTorch inference on x86 CPUs, improving median throughput by 18.43% for YOLO26n and 14.09% for YOLO11n. #25983

Fixed dataset fraction sampling that selected images from a single class, raising YOLO11n-cls top-1 accuracy from 9.89% to 87.64% in the validation run. #25968

Restored ByteTrack's low-confidence recovery step, which was unreachable under the default configuration and discarded detections in the 0.1–0.25 confidence range. #25034

Removed per-object GPU stalls from result processing, making Results.plot() 7.55× faster and cutting segmentation-validator synchronizations from 876 to 2. #25230 · #25853

Found train/test contamination affecting 27.9% of KITTI depth training images; validation found no statistically significant accuracy impact. #25650

Fixed a float-equality bug that degraded steady-frame-rate tracking by as much as 4.86 HOTA and 236 extra ID switches on DanceTrack. #531

Fixed the CLI, demo, and benchmark paths silently assuming 30 fps—a default that was wrong for every SportsMOT and DanceTrack sequence supported by the benchmark. #569

Fixed OC-SORT omitting low-confidence inputs from its return value, restoring one output row per detection. #566

Reduced a Kalman filter noise-calculation hot path by 64.93% while preserving bit-identical output. #572

Fixed spike amplitudes being scaled with channel gains as offsets, an error of +6,390 µV that flipped the sign of every spike on Intan recordings. #4765

Made amplitude scaling 2.41× faster on a 384-channel analyzer (9.23× on real Spike2 data), with bit-identical output. #4764 · #4786

Fixed synchrony metrics crediting the wrong units on multi-segment recordings while making them 8.84× faster and cutting peak memory from 1,052 to 316 MiB. #4770

Parallelized bad-channel detection, which now runs 4.30× faster on real 384-channel Neuropixels traces with identical labels. #4750