A developer has converted 11 model families to Apple's Core AI framework, making them runnable on-device on iPhones and Macs without internet dependency.

What happened: A developer ported language models and vision models (including Qwen, GLM, Gemma, and others) to Apple's Core AI framework (.aimodel format) for iOS 27 / macOS 27. The converted models are available for download with their source code, verification notes, and compression guides. The work succeeds an earlier CoreML-Models project and includes text models, vision-language models, and object detection models.

Why it matters: Apple's on-device framework lets these models run directly on a user's phone or Mac without sending data to external servers. For businesses and individuals handling sensitive information, this removes the dependency on cloud inference services. The models cover a range of sizes—from 0.8B parameters on iPhone to 35B-parameter variants on Mac—so developers can choose based on their device constraints.

What to watch: Performance varies significantly by device and model size. On iPhone 17 Pro's GPU, smaller models like Qwen3.5-0.8B achieve 71.9 tokens per second, while larger models like Qwen3.6-27B on M4 Max reach 15.9 tokens per second. The repository includes detection and instance segmentation models achieving 33–39 FPS live on iPhone 17 Pro. Code, model cards, and guides for conversion, compression, and custom optimization are available under open licenses.