OnChip Lab: Inference From the Metal Up

Field notes

The documented workflow centers on loading a local model into an OnChip runtime and generating a stream through platform bindings, with the same behavior intended across supported environments. The page names six SDK targets—Swift, Kotlin, React Native, Flutter, TypeScript, and C++—and separates silicon-specific engines from the shared runtime. It also describes LLM, VLM, speech-to-text, text-to-speech, and embedding workloads, although the supplied evidence does not verify the claimed engines or benchmarks.