Summary: On January 14, 2026, Ultralytics put YOLO26 into general availability through the Ultralytics Platform and its Python package. Dual licensing carries over: AGPL-3.0 on the open side, or an Ultralytics enterprise license bundling production support, long-term maintenance, and edge deployments that scale. Founder and CEO Glenn Jocher had given the model family its first public preview at YOLO Vision 2025 in London, and the release now comes in five scales: nano, small, medium, large, and extra-large.
On January 14, 2026, Ultralytics put YOLO26 into general availability through the Ultralytics Platform and its Python package. Founder and CEO Glenn Jocher had given the model family its first public preview at YOLO Vision 2025 in London, and the release now comes in five scales: nano, small, medium, large, and extra-large. Parameter counts were not the point of this upgrade; removing deployment friction was. Concretely: inference is end-to-end and NMS-free natively, the detection head sheds DFL and grows lighter, and the nano variant’s ONNX inference on standard CPUs runs up to 43% faster than YOLO11n on identical Xeon-class hardware.
Over the same week that rival frontier labs kept outdoing one another with bigger models, YOLO26 moved the opposite direction: a real-time vision model sized so drones, smart cameras, and embedded platforms can genuinely absorb it, with the five-size ladder climbing to production servers. Ultralytics positions the release as a fresh standard for edge-first Vision AI, and Glenn Jocher — founder and CEO — described the hard part as “making sure users can get the most out of YOLO26 while still delivering top performance.”
Removing NMS: The End-to-End Wager
Within object detection, Non-Maximum Suppression is the conventional post-processing stage, and it is also the component most likely to fail when a model is exported and deployed. YOLO26 brings in a one-to-one detection head that can produce end-to-end predictions with NMS switched off, capped at 300 boxes per image and with a clean output shape. The familiar one-to-many head stays available, and training updates both heads no matter which one you serve. The documentation states the price plainly: the end-to-end head scores roughly 1 to 2.3 AP under its non-end-to-end counterpart. It is a deliberate swap of a little accuracy for engineering convenience, and the engineers who own the deployment are the ones who make the call. In its engineering write-up, Ultralytics argues the payoff from the deployment side: NMS translates awkwardly into fixed-format export graphs, so cutting it eases the job of shipping a single model onto edge runtimes — drones, smart cameras, and embedded boards — where the glue code around post-processing is usually the first thing to break.
By the Numbers: 2.4M to 55.7M Parameters
The paper cited in the official documentation puts YOLO26 between 40.9 and 57.5 mAP on COCO detection across the five scales, at a T4 TensorRT latency of 1.7 to 11.8 ms. At the small end, yolo26n carries 2.4M parameters; at the large end, yolo26x carries 55.7M. Measured against YOLO11, the documented gains reach +2.5 box AP in detection, +3.7 mask AP in instance segmentation, +7.2 AP in pose estimation, and +3.4 mAP for oriented bounding boxes on DOTA-v1.0. The open-vocabulary sibling YOLOE-26x arrives at 40.6 AP on LVIS when driven by text prompts.
An Edge-Oriented Deployment Checklist
The task roster is wide: detection, instance and semantic segmentation, depth estimation, classification, pose, and oriented bounding boxes, each supported across training, validation, inference, and export. Export targets cover TensorRT, ONNX, CoreML, LiteRT, and OpenVINO, and exporting fuses away the auxiliary branches that only training needs, trimming the artifact further. With DFL gone, the detection head turns simpler and the export path cleaner — an immediate gain for teams fitting models into mobile apps or drones. One caveat for planning: the P2 (small-object) and P6 (large-input) variants exist only as YAML architectures for now, with no released weights, so choosing them means running the training yourself.
A Training Recipe With Three Adjustments
MuSGD combines Muon and SGD within a single optimizer, transplanting techniques from the LLM training world into vision. ProgLoss (Progressive Loss Balancing) walks the supervision signal over, step by step, to the head that actually runs at inference time. STAL (Small-Target-Aware Label Assignment) raises positive-label coverage for small objects, which matters in scenes with distant or tiny targets. Senior machine learning engineer Jing Qiu described the working style as refusing to hurry — “no rush” — with the effort going solely into refining the balance between speed and accuracy.
Licensing and the Ecosystem Reality Check
Dual licensing carries over: AGPL-3.0 on the open side, or an Ultralytics enterprise license bundling production support, long-term maintenance, and edge deployments that scale. Closed-source commercial products need to price that license in from the outset. Teams already working in the Ultralytics Python workflow receive the model as a routine package update; the genuine migration cost lies in re-benchmarking the NMS-free path against the current pipeline before switching weights. The wider picture: YOLO26 is not chasing headlines against flagship multimodal systems. Its promise runs elsewhere — 2.4M parameters, CPU-class hardware, one command to export onto nearly any runtime — and within the model race that opened 2026, this quiet edge-side line of progress earns a place on the watchlist as well.