Model requirements

What each model needs

Two numbers decide whether a model runs on your device: the memory budget it needs, and the space its container takes on the drive. Minirun checks both before anything starts. If the budget a model needs is larger than the device can safely give, the app names the constraint instead of starting the run and failing part-way through it.

Memory here means the budget the run owns, not the memory in the device. Minirun keeps headroom for the operating system and the interface, so the largest budget the app offers you is always below the figure printed on the specification sheet.

How long one reply can be

A conversation turn has a ceiling on the number of new tokens it may produce. It is a working-set bound rather than a speed setting, and a reply that stops at the ceiling has finished rather than failed.

Kimi K3 and DeepSeek V4 Flash both answer up to 64 new tokens per turn on a Mac, and DeepSeek V4 Flash does the same on iPhone. Kimi K3 on iPhone is the exception: it answers a couple of words per turn today, which is enough to see a 1.56 TB model produce text on a phone. Longer replies there are coming.

What the three presets do

The preset chips are not quality levels. Each one is a point at which the plan for the run changes — the budget at which one more part of the model can be afforded a permanent place in memory instead of being read for every token.

• A preset your device cannot afford is shown disabled, with the amount it is short by. It is never hidden and never quietly collapses onto the smallest one.
• A budget below the model's minimum is declined by name. It is never raised for you to something that would have worked.
• More memory helps in a model-specific way. A model with no runtime in the app has no dial at all.

Floor

The smallest budget the model is allowed. Nothing stays resident: every layer is read from the drive for every token. This is what a conversation that never touched the dial uses.

Balanced

Enough headroom to keep the parts that pay for themselves the most. On a Mac that can reach it, DeepSeek V4 Flash keeps all of its deterministic layers and its output head at this setting.

Generous

The largest budget your device can offer. For DeepSeek V4 Flash that is the device ceiling; for Kimi K3 it is the highest setting that still leaves room for the reads the run stages ahead of itself.

The budget changes what a run costs, not what it says

The dial decides how much of the model stays in memory and how much is read as the run reaches it. It does not change the answer. The same prompt on the same device produces the same text whether every layer streams from the drive or every layer is held resident — what changes is the time each token takes and how much memory the run occupies.

That holds within one device. Two different devices are not promised the same words: the same question answered on an iPhone can word the end of a sentence differently from the Mac.

How long the first check takes

Before a container can be used in a chat, Minirun reads every byte of it once and checks it. The time that takes follows the size of the container and the speed of the link, not the memory budget. A complete Kimi K3 check on an iPhone 16 Pro, with the container on an external NVMe drive on the phone's own port, takes around half an hour.

Keep the device powered and the drive mounted for the whole pass. A cancelled pass leaves the copy unchecked and has to be started again.