CVNets is Apple’s open-source library for training computer vision networks, covering image classification, object detection, semantic segmentation and foundation-model training such as CLIP. Developed by Apple AI/ML research, it provides clean, reproducible implementations of the team’s architectures (including the MobileViT family) alongside standard backbones, with training recipes for benchmarks like ImageNet, MS-COCO, ADE20K and Pascal VOC. The repository has around 2,000 GitHub stars and is actively maintained, with updates in 2026.
CVNets is designed as a research-grade training toolkit rather than an inference-only model zoo. A single configuration system drives data loading, augmentation, model construction, distributed training and evaluation, so a recipe that trains a mobile-sized classifier can be adapted to detection or segmentation with modest changes. The library emphasizes mobile and efficient architectures, reflecting Apple’s on-device focus, while still supporting large vision transformers and multi-modal CLIP-style training.
Free. CVNets is openly released by Apple for research and development use; you train on your own hardware.
Computer vision researchers and engineers who need a solid training framework for vision models, especially those targeting efficient on-device deployment.
CVNets is the training codebase behind several of Apple’s published vision papers, and its model zoo includes the MobileOne and MobileViT variants that are popular starting points for on-device vision work. The toolkit uses a modular design in which losses, optimizers, data transforms and model blocks are interchangeable components, so researchers can assemble a new architecture or ablation study from configuration files instead of forking training loops.