Run models larger than memory.

Chat with open models bigger than your device’s memory. Weights stream from your SSD through a memory budget you set. Free, local, on Mac and iPhone.

Download for Mac Get it for iPhone

Mac: macOS 15 or later on Apple silicon, and it updates itself. iPhone: through TestFlight, iPhone 15 Pro or later on iOS 18.

The Minirun Mac app with a finished DeepSeek V4 Flash answer on screen and the Instruments panel beside it, showing the decode rate, the data read per token and the footprint against the budget.An iPhone 16 Pro running Minirun, showing the instruments for a completed Kimi K3 turn: the time per token, the data read for it, and the storage flow behind it.

Bigger than memory.

Kimi K3 is 1.56 TB and runs on an iPhone 16 Pro. DeepSeek V4 Flash, 167 GB, chats on a Mac. The model stays on your SSD.

Browse all models
Chat settings for Kimi K3 on an iPhone 16 Pro: a 5.80 GB memory budget set against a 6.44 GB device limit, with the data the run reads for each token listed beneath it.

Set a memory budget.

Pick Floor, Balanced or Generous, or drag the dial. Minirun holds what fits, streams the rest, and never crosses the number you chose.

The Minirun chat settings panel on macOS: the three budget presets, the chosen budget, the dial against the device limit, and what the runner holds resident.

Runs from your SSD.

Add a folder on an external NVMe drive. Minirun finds the models in it and streams their weights straight off it, on Mac and iPhone.

How storage works
The Minirun storage pane on macOS: a granted folder on an external NVMe volume, with the sequential and scattered read rates it was assessed at and the space left on it.The same storage pane in the Minirun iPhone app, showing the granted folder on the external drive the phone reads its models from.

Watch every token.

Instruments shows the decode rate, the data read per token, the footprint against your budget, and the flow from storage, live while the model answers.

The Minirun instruments panel on macOS during a run: the decode rate, the data read for the current token, the footprint bar against the budget, and the storage data flow chart.

One library, both devices.

Models you keep on the drive show up in the app on Mac and iPhone alike. Find new ones from Hugging Face without leaving the app.

The Minirun models list on macOS showing Kimi K3 at 1.56 TB and DeepSeek V4 Flash at 167 GB, each sitting on the drive it was found on.The same library on an iPhone 16 Pro: Kimi K3 at 1.56 TB and DeepSeek V4 Flash at 167 GB, both found on the drive attached to the phone.

Everything stays on your device.

Prompts, answers and weights never leave your Mac or iPhone.

Free and open.

Both apps are free, and the source is on GitHub.

Always current.

The Mac app updates itself; new models appear in the app’s Find Models.

Get Minirun.

Mac

Apple silicon. Downloads as a disk image and updates itself.

Download for Mac

macOS 15 or later · 11 MB

iPhone

Through TestFlight. Free, no account beyond your Apple ID.

Get it for iPhone

iOS 18 or later · iPhone 15 Pro or later

Quick start