What Minirun is
Minirun is an app for Mac and iPhone that runs open models locally — including models far larger than the memory in the device. Kimi K3 is 1.56 TB and DeepSeek V4 Flash is 167 GB, and both hold a conversation on Apple silicon, because Minirun streams their weights from your SSD as the run needs them instead of loading the whole model into memory first.
You set a memory budget — the dial in Chat settings — and the engine plans the entire run inside it: which regions stay resident, which are read as they are reached, and when each one is released. If the run cannot be represented inside the budget, Minirun says so before it starts rather than halfway through.
Nothing leaves the device to answer you. Minirun uses the network to find and download containers, and for nothing else.
Core components
The engine has three cooperating components. Each has one responsibility and one dedicated reference page.
- Minirun Container
- A complete, addressable description of the model regions the runtime may read.
- SSD Streaming
- Planned reads from internal or external storage, overlapped with compute when the model allows it.
- Memory Dial
- The memory budget, set in Chat settings. It is the ceiling used for admission, residency, read-ahead, and model-specific cache decisions.
Capability pipeline
Discovery, local storage, verification, runtime support, and device admission are separate capabilities. A model can pass one stage without passing the next; the interface reports the last established state instead of collapsing everything into an Installed badge.
Local by default
Prompts, generated text, and conversation history stay on the device. Minirun uses the network to discover repositories and transfer model containers from Hugging Face; a supported conversation does not send its prompt to a hosted inference service.