# Minirun

A native macOS and iOS app that runs open-weight language models far larger than the device's memory by streaming each layer's weights from an external SSD within a memory budget the user sets, entirely on-device.

Minirun runs open models whose weights are far larger than a Mac or iPhone's available memory. Rather than loading a model fully into memory, it reads each layer, and for mixture-of-experts models only the experts the router selects, from an external NVMe drive into a small fixed set of buffers, hands them to MLX without copying, and releases them, all within a memory budget the user chooses ahead of time.

The app refuses to exceed the chosen budget rather than running out of memory mid-answer, and shows live instruments such as decode rate and bytes read per token while a model answers. Model weights are distributed as byte-preserving repacks of published checkpoints, verified against the published Hugging Face file tree before use, and prompts, answers and telemetry never leave the device. It runs on both macOS and iOS on Apple Silicon.

Website: https://minirun.dev/