ModelsGitHub

Quick start

Go from a fresh installation to a verified local conversation.

Find or add a model

Open Settings → Models. The main list is your local library: containers found in the folders you have granted. Choose Find Models to open the separate Hugging Face catalog when you want a new model.

  1. Use an existing copy

    Open Storage, add the folder that contains the Minirun container, and wait for the scan to complete.

  2. Download a copy

    Open Find Models, select a published container, choose Download, and confirm a writable destination through the system picker.

  3. Return to the local library

    After the container is present, open its local copy from Models.

Verify before chat

Open the local copy and choose Verify all files. A spot-check does not enable chat. Full verification is a bounded streaming operation, so it does not materialize an entire weight file in memory.

Create the conversation

  1. Choose New chat

    The model picker shows only fully verified models with a compatible chat runtime on this device.

  2. Set a memory budget

    Use a preset or the slider. The selected value must be at or above the model's displayed minimum.

  3. Send a message

    Minirun assembles the model's supported chat format, builds a memory plan, and starts the local run.

  4. Open Instruments when useful

    Inspect stage, footprint, reads, overlap, and model-specific signals reported by the active runtime.

Expected result

The conversation picker lists the verified local runtimes that are available on this device. If the expected model is absent, use the troubleshooting decision path rather than repeating a download or verification without a named reason.

Two things surprise people on a first run, and the app states both rather than hiding them. Every model has a minimum budget: below it Minirun declines the run and names the constraint instead of quietly raising the number for you. Every model also has a limit on how long one reply may be, and on iPhone Kimi K3 answers a couple of words per turn — a short reply there is the published limit, not a fault. The check is the long part: a complete Kimi K3 pass on a phone takes around half an hour.