CoreNet is Apple’s open-source library for training deep neural networks, covering everything from image classification and object detection to segmentation and vision-language foundation models. It has around 7,000 stars on GitHub and is actively maintained in 2026.
CoreNet grew out of Apple’s CVNets project into a general-purpose training toolkit used by Apple’s own research teams. It provides modular building blocks and complete, reproducible training recipes for a wide range of tasks: image classification, object detection, semantic segmentation, and multimodal models in the style of CLIP — including work connected to Apple’s MobileCLIP family of efficient vision-language models. The library emphasizes efficient architectures suitable for on-device deployment, configurable training pipelines, and support for standard benchmarks and datasets. Because it comes from the team that ships models onto billions of devices, its recipes reflect hard-won practical experience with training efficient networks. The library covers a broad catalog of architectures — from ResNets and MobileNets to Vision Transformers and hybrid designs — together with knowledge-distillation options, standard augmentation pipelines and evaluation harnesses, so a recipe can be adapted to a new dataset or model family with modest configuration changes. CoreNet is free and open source on GitHub under Apple’s license terms.
CoreNet is free and open source. Apple publishes the code and training recipes on GitHub at no cost; you only need your own compute for training runs.
CoreNet is for computer vision researchers and engineers who want production-informed training recipes from Apple’s teams, especially those building efficient models destined for phones and other edge devices.