Languagemodelsession

A technical, data-driven walk-through optimizing a handwritten matrix multiply for Swift-based LLM training on Apple Silicon, showing stepwise gains from plain C parity to AMX and tiled Metal with multi-threading, culminating in a claimed 1.1+ TFLOP/s throughput and ~11 tokens/s on TinyShakespeare.

On-device AI should be the default to improve privacy, reliability, and developer control, demonstrated with Apple tooling and a Brutalist Report on-device summarization example.

Apfel exposes Apple’s on-device LLM on macOS 26+ as a three-front interface (CLI, OpenAI-compatible HTTP server, and interactive chat) with zero external cost, no API keys, and full local inference.