Awesome Core AI

Curated resources for Apple's Core AI framework (iOS 27 / macOS 27+) — official tooling, converted models, conversion pipelines, sample apps, benchmarks, and learning material.

Core AI is Apple's next-generation on-device ML stack (successor to Core ML): models ship as .aimodel bundles and run on Apple silicon GPU / Neural Engine. This list tracks the ecosystem growing around it.

PRs welcome — see Contributing.

---

Contents

• Official
• Getting started
• Running models in your app
• Models
• Conversion
• Serving
• Benchmarks & engineering notes
• Learning

Official

• apple/coreai-models — The reference repo: model export recipes, Python primitives, and the Swift runtime packages (CoreAILM, CoreAIObjectDetection, CoreAISegmentation, CoreAISpeech, CoreAIDiffusion).
• apple/coreai-torch — PyTorch → Core AI IR: torch.export conversion, composite ops, custom op lowering, inline Metal kernels.
• apple/coreai-optimization — Quantization, palettization, and compression for Core AI deployment.
• WWDC26 sessions 324 · 325 · 326 · 330 — Core AI introduction through advanced topics (session 330 covers quantized TensorOps matmul and FlashAttention).

Getting started

• coreai-kit ChatDemo — Open the Xcode project, pick a model in the picker, chat fully on-device; the model downloads once and is cached. Also a terminal path: swift run chat-cli.
• timokoethe/CoreAIChat — Minimal SwiftUI chat app for macOS showing the smallest CoreAILanguageModel → LanguageModelSession wiring.
• rbniranjan/WWDC2026CoreAI — Hands-on examples following the WWDC26 sessions.
• The Art of Core AI — Free hands-on book: 13 chapters + labs from first export to custom kernels (source, Japanese edition on Zenn).

Running models in your app

• john-rocky/coreai-kit — One line of Swift per model (ChatSession(catalog: "qwen3.5-2b")) across LLM / VLM / ASR / TTS / diarization / detection and more; 49-model pinned catalog, nightly on-device gate. FoundationModels-compatible: catalog models plug into LanguageModelSession with tool calling and guided generation. SPM.
• rudrankriyam/Core-AI-Framework-Lab — Practical lab app: model asset management, specialization states, compute-unit configuration, benchmarking across modalities.
• Techopolis/AFM-Studio — Chat app spanning Apple Foundation Models, Private Cloud Compute, and Core AI models behind one provider interface.
• mweinbach/NemotronCoreAI — SwiftPM streaming-ASR runtime for NVIDIA Nemotron 3.5 on Core AI.

Models

• john-rocky/coreai-model-zoo — 49 models converted and device-verified (iPhone 17 Pro / M4 Max): LLM, VLM, OCR, ASR, TTS, diarization, image/video/music generation, forecasting. Each with a downloadable Hugging Face bundle, conversion recipe, and a ready-to-build app.
• Hugging Face: mlboydaisuke — The zoo's published .aimodel bundles.
• SAL2-Dev/ComfyUI-CoreAI — Core AI vision nodes (depth, detection, VLM, CLIP, on-device LLM) for ComfyUI.
• kevinqz/coreai-catalog — Source-grounded registry of Core AI models, artifacts, upstreams, and provenance.
• tmorales2000/fastvlm-coreai — FastVLM vision-language model converted to Apple Core AI .aimodel format for Neural Engine deployment on iOS/macOS 27.
• 0Itsuki0/Swift_STTWithCoreAI — A demo of using the CoreAI framework (OS 27+) for on-device Speech to text.

Conversion

• apple/coreai-torch — The official bring-your-own-PyTorch-model path (see Official).
• devin-lai/coreai-onnx — Convert ONNX models directly to .aimodel.
• coreai-model-zoo/conversion — Reproducible per-model conversion recipes (pinned base + overlay) behind every zoo bundle, with a doctor/run CLI.
• lucasnewman/mlx2coreai — Convert MLX models to Core AI.
• weichao1221/coreai-models-gui — SwiftUI GUI for Core AI model export, with ModelScope source support.
• NagaYu/silicon-forge — Pull, convert, quantize, benchmark, and package open-weight LLMs for Apple silicon (MLX → Core AI).
• massif-01/qwen3-1.7b-coreai-ios — High-fidelity Qwen3-1.7B W8 onboarding with A17 Pro Neural Engine validation for Apple Core AI.

Serving

• RedHillsMediaFL/caix — Native Core AI inference server for Apple silicon: OpenAI/Anthropic-compatible API, dashboard, streaming chat.

Benchmarks & engineering notes

• coreai-model-zoo decode-throughput tables — Device-verified tok/s (greedy, top-1 exact vs the Hugging Face reference) on iPhone 17 Pro GPU/ANE and M4 Max.
• coreai-model-zoo/knowledge — Port write-ups and engineering notes: ANE vs GPU trade-offs, quantization ladders, chunked prefill, speculative decoding, custom Metal kernels.
• rwrun/coreAIvsMLLLM — Core AI vs MLX LLM comparison on iOS (Russian).

Learning

• The Art of Core AI — see Getting started.
• WWDC26 sessions — see Official.
• Steven-ZN/Apple-26-ai-skill — Guide coding agents to instrument, trace, compare, and optimize Apple Foundation Models, PCC, Core AI, and multi-session agentic flows using Instruments, str...

Contributing

PRs welcome. Criteria for inclusion:

• Public repo (or published resource) that is specifically about Apple's Core AI framework / .aimodel.
• Has a README that lets a stranger use it: what it is, how to run it, what OS/hardware it needs.
• Model entries should state the license and how correctness was verified (e.g. parity vs the upstream reference).

One line per entry, factual tone, no superlatives. Within each section: official Apple resources first, then entries ordered by how useful and proven they are for that section's purpose — never by authorship. New radar finds join at the bottom until they earn a higher spot.

New entries are also scouted and added weekly by an automated radar (GitHub / Hugging Face search): confident finds — a strong Core AI signal plus some traction — land in the list automatically, and everything else waits in RADAR.md until it qualifies. Spotted a bad entry? Open an issue or PR; removals are pinned in .github/radar-ignore.txt so the radar never re-adds them.