Own your AI. The native macOS harness for AI agents -- any model, persistent memory, autonomous execution, cryptographic identity. Built in Swift. Fully offline. Open source.

🔎 SimilaritySearchKit is a Swift package providing on-device text embeddings and semantic search functionality for iOS and macOS applications.

Train and run transformers directly on Apple's Neural Engine in Swift bypass coreml entirely

Your models on any xPU

PyTorch → CoreML conversion pipeline for Kokoro TTS. Unlocks fast on-device text-to-speech on Apple Neural Engine.

ModernBERT model optimized for Apple Neural Engine.

macOS meeting transcription app with speaker diarization

Apple Neural Engine (ANE) LLM inference engine — reverse-engineered private APIs, Metal GPU shaders, hybrid ANE+GPU+CPU on Apple Silicon. 32 tok/s matching llama.cpp, 3.6 TFLOPS fused ANE mega-kernels.

Push-to-talk voice dictation for macOS. 100% local, free, open source. Apple Silicon MLX. No cloud, no subscription.

Apple FoundationModels API on iOS 18+. Same call site, native passthrough on iOS 26 (Apple Intelligence), CoreML / MLX backends on older OSes. Drop-in source compatible.

Push-to-talk dictation for macOS Apple Silicon. On-device speech recognition via Parakeet TDT v3 on the Apple Neural Engine (ANE) + CoreML. ~100 ms from key release to pasted text, 2.2 MB download, ~80 MB RAM. Native Swift, no cloud.

Train transformers on Apple's Neural Engine. Autonomous hyperparameter search via Karpathy's autoresearch protocol. 43 experiments, 8 verified findings.

First super-resolution model designed for Apple Neural Engine. 2x upscale, real-time, on-device. Built by Ben Racicot.

First speculative decoding using Apple Neural Engine as draft + GPU as verifier on Apple Silicon. Custom minGRU encoder for heterogeneous compute.

A toolkit for BLE digital stethoscopes — capture, DSP, log-mel spectrograms, and murmur classification on the Apple Neural Engine

Reverse-engineering field guide for Apple Neural Engine (ANE) — hardware constraints, IOSurface layout, MIL programming, and undocumented behavior.

Research notes on training a custom gated MLP directly on the Apple Neural Engine. Architecture decisions, results, and observations from an on-device training prototype.

Neural network that generates structured design specifications from type embeddings using ANE-accelerated training

Direct ANE kernel execution for speculative decoding on Apple Silicon — Qwen3-0.6B on Neural Engine + Qwen3.5-9B GPU verify

LLM training on Apple's Neural Engine — native Obj-C, private APIs, zero GPU. Dynamic weight pipeline for training without kernel recompilation.