400+ open models on Hugging Face, converted to LiteRT, Core AI and ExecuTorch and measured on phones. 163 of them in litert-community. PRs merged in litert-samples (32) and executorch (9). I write as MLBoy on Zenn, Medium and Qiita.
For local chat, speech or image understanding in a Swift app, start here to try the Mac app, choose a model and find package examples. CoreAIKit Skills helps your coding agent resolve model IDs, release pins and device variants before adding the feature.
coreai-model-zoo — downloadable Core AI models with recipes and validation records · CoreAIKit — Swift APIs for local models · devicemark — on-device LLM leaderboard for iPhone · The Art of Core AI — the textbook (JA, free; EN)
CoreML-Models ⭐1.8k · LiteRT-Models · VLMKit · ex-Ultralytics (YOLO on mobile) · 10 apps on the App Store · Osaka, Japan (JST)
Models: Hugging Face · litert-community · coreai-community
| Ship a model on-device | PyTorch / ONNX → Core ML or LiteRT, production-ready, not a demo |
| Make it fast | Latency and memory — quantization, palettization, ANE / GPU / NPU tuning |
| Real-time camera & video | Detection, tracking, segmentation at frame rate |
| On-device VLM / LLM | Structured extraction and generation, fully offline |
| Train or fine-tune | Building the model for the target device, not only deploying someone else's weights |
Small, well-scoped pieces are welcome as well as longer engagements.
📧 rockyshikoku@gmail.com — I reply within a day 🌐 john-rocky.github.io — shipped apps, past work, details
Freelance through Pebble Inc. (Osaka, JST) — currently taking projects. Remote; occasional travel is fine.