On January 14, 2026, Ultralytics released YOLO26 to the public through the Ultralytics Platform and its Python package. The model family had been first previewed by founder and CEO Glenn Jocher at YOLO Vision 2025 in London, and it now ships in five scales — nano, small, medium, large, and extra-large. The upgrade is not about stacking parameters; it is about removing deployment friction: native NMS-free end-to-end inference, a lighter detection head with DFL stripped out, and ONNX inference on standard CPUs up to 43% faster for the nano variant than YOLO11n on the same Xeon-class hardware.
While frontier labs spent the same week shipping larger models, YOLO26 heads the other way: a real-time vision model that drones, smart cameras, and embedded platforms can actually absorb, scaling up through five sizes to production servers. The company frames it as a new standard for edge-first Vision AI, and founder and CEO Glenn Jocher described the hard part as “making sure users can get the most out of YOLO26 while still delivering top performance.”
Dropping NMS: The End-to-End Bet
Non-Maximum Suppression is the standard post-processing step in object detection — and the piece most likely to break during export and deployment. YOLO26 introduces a one-to-one detection head that runs end-to-end predictions directly with NMS disabled, emitting up to 300 boxes per image with a clean output shape. The conventional one-to-many head remains available, and both heads are trained regardless of which one you serve. The cost is documented plainly: the end-to-end head scores roughly 1 to 2.3 AP below the non-end-to-end path. That is an explicit trade of a little accuracy for engineering convenience, and the engineers responsible for deployment get to make the call themselves. Ultralytics’ own engineering write-up frames the payoff in deployment terms: NMS is awkward to express in fixed-format export graphs, so removing it strips friction from shipping the same model onto edge runtimes — drones, smart cameras, and embedded boards — where post-processing glue code tends to break first.
The Numbers: 2.4M to 55.7M Parameters
Per the paper cited in the official documentation, YOLO26 spans 40.9 to 57.5 mAP on COCO detection across its five scales, at 1.7 to 11.8 ms T4 TensorRT latency. The smallest yolo26n sits at 2.4M parameters; the largest yolo26x at 55.7M. Against YOLO11, the reported gains run up to +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, reaches 40.6 AP on LVIS using text prompts.
A Deployment List Built for the Edge
Task coverage is broad: detection, instance and semantic segmentation, depth estimation, classification, pose, and oriented bounding boxes — all with train, validation, inference, and export support. Export targets include TensorRT, ONNX, CoreML, LiteRT, and OpenVINO, and fused exports drop the training-only auxiliary branches to shrink the artifact further. Removing DFL simplifies the detection head and cleans up the export path, which pays off directly for teams squeezing models into mobile apps or drones. One caveat for planners: the P2 (small-object) and P6 (large-input) variants ship as YAML architectures only, without released weights, so those configurations mean training your own.
Three Changes in the Training Recipe
MuSGD is a hybrid Muon-plus-SGD optimizer that ports techniques from the LLM training world into vision. ProgLoss (Progressive Loss Balancing) gradually shifts supervision toward the head that actually runs at inference time. STAL (Small-Target-Aware Label Assignment) improves positive-label coverage for small objects, which matters for distant or small-target scenes. Senior machine learning engineer Jing Qiu described the build philosophy as staying steady — “no rush” — refining only the speed-accuracy balance.
Licensing and the Ecosystem Reality
YOLO26 keeps dual licensing: AGPL-3.0 for open use, or an Ultralytics enterprise license that bundles production support, long-term maintenance, and scalable edge rollouts. Closed-source commercial products have to price that license in from day one. For teams already inside the Ultralytics Python workflow, the model arrives as an ordinary package update; the real migration cost sits in re-benchmarking the NMS-free path against the existing pipeline before swapping weights. The bigger picture: YOLO26 is not competing with flagship multimodal systems for headlines. It pushes a different promise — 2.4M parameters, CPU-class hardware, one command to export onto almost any runtime — and inside the model race that opened 2026, this quiet edge-side line of progress deserves a spot on the radar too.
Sources
- Ultralytics Redefines State-of-the-Art Vision AI with YOLO26 — Ultralytics
- Ultralytics YOLO26 Documentation — Ultralytics
- Why Ultralytics YOLO26 Removes NMS and How That Changes Deployment — Ultralytics Blog
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.