Super-Gradients

Super-Gradients is an open-source training library from Deci (now part of NVIDIA) that makes it dramatically easier to train and fine-tune state-of-the-art computer vision models in PyTorch. With more than 5,000 stars on GitHub and development continuing into 2026, it is best known as the home of YOLO-NAS, and it bundles proven training recipes so you can reproduce strong results without weeks of hyperparameter tuning.

Overview

Training modern vision models well requires the right combination of augmentations, losses, optimizers, EMA, mixed precision and scheduling. Super-Gradients packages all of that into ready-made recipes for classification, object detection, pose estimation and semantic segmentation. You pick a model from the built-in zoo, point it at your dataset, and train with settings that the Deci team has already validated, then export for deployment.

Key Features

• Production-ready training recipes for classification, detection, segmentation and pose estimation
• Model zoo including YOLO-NAS, PP-LiteSeg, ResNet, EfficientNet and ViT variants with pretrained weights
• Built-in support for mixed precision, distributed training, EMA and QAT (quantization-aware training)
• One-line dataset handling for COCO, Pascal VOC, Cityscapes and custom formats
• Export paths to ONNX and TensorRT for deployment
• Automatic mixed precision, gradient accumulation and checkpoint management included in every recipe
• Extensive documentation, notebooks and a reproducible benchmark suite

💳 Pricing

Free and open source under the Apache 2.0 license. There is no paid tier for the library itself; you train on your own hardware or cloud GPUs.

Who Is It For

Computer vision engineers and ML teams who want reliable, reproducible training for production vision models without assembling and tuning a training stack from scratch. It is a particularly good fit when a team needs detection or segmentation models that must later run at the edge, since the same recipes support quantization-aware training and deployment export.