Ultralytics Rebuilds YOLO26 to Hit 57.5 AP With Zero Post-Processing
Ultralytics YOLO26 drops NMS post-processing, removes Distribution Focal Loss, and borrows an optimizer trick from LLM training to push real-time object detection to 57.5 mAP on COCO at under 12 ms latency.
Takeaways
- NMS-free by default: YOLO26 uses a dual-head design where the default one-to-one head produces up to 300 detections with no post-processing required. [Paper]
- DFL removed: Distribution Focal Loss is dropped entirely, cutting 12% of parameters and 20% of FLOPs in the nano model, and improving large-object AP at high resolution.
- MuSGD optimizer: Adapted from LLM training, MuSGD reaches +0.4 mAP in 500 epochs vs. 600 epochs for standard SGD, a 17% training cost reduction.
- STAL fixes small-object blindness: Tiny objects below the minimum feature-map stride are temporarily inflated during label assignment so they always receive gradient signal.
- COCO benchmarks: 40.9-57.5 mAP at 1.7-11.8 ms T4 TensorRT latency; nano model runs 43% faster on CPU than YOLO11n. [Docs]
- Open-vocabulary extension YOLOE-26x reaches 40.6 AP on LVIS minival under text prompting, beating DetCLIP-T by +6.2 AP. [Code]
Ultralytics YOLO26 is the most significant overhaul of the YOLO architecture in years. Rather than the usual incremental accuracy bump, this release rethinks three core pain points that have quietly taxed every YOLO deployment: a post-processing step that adds latency and complicates export, a bounding-box regression head that bloats small models, and a training recipe that leaves tiny objects invisible to the optimizer. The result is a model family that is simultaneously lighter, faster to train, and more accurate than its predecessor.
Why the old recipe was showing its age
Every YOLO model since YOLOv8 has shared the same structural debt. Most CNN-based detectors still rely on non-maximum suppression (NMS) at inference. YOLOv10 introduced a dual-head design to remove NMS, but applies identical loss weights to both heads throughout training: the one-to-one branch, the only head used at inference, receives the same optimization pressure as the dense one-to-many branch, leaving it under-trained relative to what targeted supervision could achieve.
The second problem is the Distribution Focal Loss (DFL) module. DFL expands bounding-box regression from 4 scalars to 4K logits per spatial location (typically K=16), inflating head parameters. For instance, YOLO11n with DFL has 2.6M parameters and 6.5 GFLOPs, whereas removing DFL reduces these figures to only 2.3M parameters and 5.2 GFLOPs, a 12% parameter and 20% FLOP reduction from this single module alone. DFL also caps how large an object can be before the head can no longer represent its extent accurately.
The third issue is small-object blindness. The Task-Aligned Learning (TAL) label-assignment strategy selects candidate anchors based on geometric containment within ground-truth boxes. After downsampling, objects whose spatial extent falls below the minimum stride have no anchor inside their box, receiving zero positive assignments and zero gradient signal. In practice, tiny objects simply disappear from training.
Four coordinated fixes
YOLO26 builds on YOLO11 and adopts a dual-head design for native NMS-free inference, removes DFL entirely yielding a lighter regression head with unconstrained range, and combines three complementary training components: MuSGD, a hybrid Muon-SGD optimizer; Progressive Loss, which gradually shifts supervision toward the inference-time one-to-one head; and STAL (Small-Target-Aware Label Assignment), which guarantees positive candidate coverage for tiny objects under TAL.

Each fix is surgical and measurable:
- DFL removal: Removing DFL alone costs 0.6 AP on the YOLO11s baseline, but this gap is fully recovered by L1 supervision (+0.2), STAL (+0.2), and backbone/neck refinement (+0.2), yielding a strictly lighter and faster head at the same accuracy. At 1280-resolution, the DFL-free head actually wins by +1.3 AP because the finite bin range of DFL becomes a hard ceiling for large objects.