Is YOLO-NAS abandoned?

SuperGradients last released 3.7.1 in April 2024. LibreYOLO supports YOLO-NAS prediction, training, validation and export on current PyTorch.

Does Ultralytics maintain YOLO-NAS?

Ultralytics supports inference, validation and export. It does not support training.

Can pretrained YOLO-NAS weights be used commercially?

Deci's pretrained weights retain their upstream non-commercial terms, whichever library loads them. Training from random initialization avoids those checkpoints. We also plan to publish from-scratch COCO weights.

All articles

YOLO-NAS Is Still Maintained

August 20, 2026Xuban

YOLO-NAS is still a useful detector. Its original home, SuperGradients, last released 3.7.1 in April 2024. LibreYOLO maintains prediction, training, validation and export for it.

from libreyolo import LibreYOLO

model = LibreYOLO("LibreYOLONASs.pt")
result = model("image.jpg", save=True)
print(result.boxes.xyxy)

LibreYOLO supports S, M and L for detection, plus pose. Checkpoints download from Deci's public CDN under names like LibreYOLONASs.pt. LibreYOLO hosts none of them, and Deci's non-commercial terms still apply. SuperGradients pins PyTorch between 1.9 and 1.13; LibreYOLO requires 2.4 or newer. Detection export covers ONNX, TorchScript, OpenVINO, NCNN, TFLite, Paddle, MNN, ExecuTorch and Core AI. CoreML is not supported.

To train without a Deci checkpoint:

from libreyolo import LibreYOLONAS

model = LibreYOLONAS(None, size="s")
model.train(data="my-dataset.yaml", imgsz=640, batch=16)

We also plan to train YOLO-NAS from scratch on COCO and publish those weights.

Install with pip install libreyolo. The YOLO-NAS docs have the model details. There is a longer SuperGradients alternative article.

GitHub | Documentation