I explore how far on-device AI can be pushed on Apple devices — from 12B-class language models on iPhone to image generation on Apple Watch.
Appleデバイス上でオンデバイスAIをどこまで動かせるかを探究しています。 iPhone上の120億パラメータ級言語モデルから、Apple Watch上の画像生成まで取り組んでいます。
Gemma 4 Core ML experiments targeting an 8 GB iPhone through model splitting, sequential loading, and aggressive compression.
モデル分割、逐次ロード、積極的な圧縮を用いて、8 GBのiPhone上でGemma 4を 動作させることを目指すCore ML実験です。
Running diffusion language model experiments on iPhone with llama.cpp,
GGUF, Metal, and SwiftUI.
llama.cpp、GGUF、Metal、SwiftUIを組み合わせ、iPhone上で
拡散言語モデルを動かす実験です。
Exploring local image generation on Apple Watch with a streamed, CPU-only Core ML diffusion pipeline.
ストリーミング方式のCPU-only Core ML拡散パイプラインを用いて、 Apple Watch単体でのローカル画像生成を探究しています。
An experimental social platform that transmits semantic representations instead of raw images and reconstructs them with on-device AI.
画像そのものではなく意味表現を伝送し、オンデバイスAIで画像を再構成する 実験的なソーシャルプラットフォームです。
On-device AI · Core ML · Swift · SwiftUI · llama.cpp
Model compression · Semantic communication · AI-assisted development
オンデバイスAI · モデル圧縮 · セマンティック通信 · AI協働開発
I build experiments together with AI coding agents, then validate them through repeatable scripts, device-side instrumentation, and tests on actual hardware.
AIコーディングエージェントと協働して実験を組み立て、再現可能なスクリプト、 デバイス側の計測機構、実機テストを通じて検証しています。
Learning in public — pushing small devices beyond their expected limits.
学びと試行錯誤を公開しながら、小さなデバイスの可能性を押し広げています。


