john-rocky/coreai-model-zoo
Community model zoo for Apple Core AI (iOS/macOS 27): 49 models — LLM, VLM, OCR, ASR, TTS, image/video/music gen, forecasting — converted, verified on real devices, downloadable from Hugging Face, and runnable in one line of Swift via CoreAIKit. Plus conversion recipes, on-device benchmarks, custom Metal kernels, and a knowledge base.
What it solves
Running large language models and other AI models on Apple devices (iOS and macOS) often requires complex conversion processes. This repository provides a centralized catalog of models already converted for Apple's Core AI, along with the exact recipes used to create them, ensuring developers can easily integrate high-performance, on-device AI into their applications.
How it works
The project provides downloadable .aimodel files that are device-verified. Each model includes a recipe.toml file that documents the exact conversion process used. Users can:
- Download and run models directly using the CoreAIKit Swift package.
- Rebuild models using provided Python conversion scripts (
zoo_convert.py) to ensure reproducibility. - Verify bundles against their source models using
zoo_verify.pyto ensure integrity.
Who it’s for
- iOS/macOS Developers looking to integrate offline, on-device AI (text, vision, audio, or robotics) into their apps.
- AI Researchers wanting to see how specific model architectures (like MoE or diffusion models) perform when ported to Apple's hardware.
- Machine Learning Engineers needing reliable conversion recipes for Apple's Core AI ecosystem.
Highlights
- Diverse Model Support: Includes LLMs, vision-language models, speech-to-text, text-to-speech, and even robotics/action models.
- Reproducible Conversions: Every model is accompanied by a recipe that can be rerun to produce the same result.
- On-Device Focus: Optimized for running fully offline on iPhone and Mac hardware.
- Extensive Catalog: Features a wide variety of architectures including Qwen, Gemma, Llama, and specialized models like BitCPM (1.58-bit ternary LLM).