How to use this page
Start with the question that matches the visible product state. Follow the recovery steps in order and stop when the named state changes. Re-downloading, re-verifying, or changing the memory budget without a matching diagnosis can discard useful work without fixing the cause.
Terms such as Published, On this device, Verified, Paused, and Not reported have exact meanings. Use the status reference when the interface wording is unfamiliar.
Models and compatibility
Why does a model appear in Find Models but not in my local library?
Find Models is the remote Hugging Face catalog. The main Models list contains only containers discovered inside storage locations registered on this device. A remote publication is not a local installation.
- Open the publication in Find Models and choose Download, or add the folder that already contains its Minirun container.
- Wait for the local storage scan to finish, then return to Models.
Why is a local model missing from the chat picker?
The chat picker is intentionally narrower than the local library. It requires a mounted local copy, matching full verification, a chat runtime for that model family, and admission on the current device.
- Open the local model detail and read its verification and support states.
- If it is Unverified, run Verify all files. If it is unsupported, verification or reinstallation will not add a missing runtime.
Why did a local model return to Unverified?
Saved verification remains valid only while the publication, complete file set, metadata, payload identities, and bound filesystem objects still match. A changed publication or local file, repaired transfer, replaced directory, or stale storage authority can invalidate the previous result. Relaunching the app by itself should not.
- Confirm the expected storage location is mounted and registered.
- Open the local copy and run Verify all files. If the pass names a mismatch, repair or replace the affected copy before retrying chat.
Why can Minirun manage a model but not chat with it?
Container management and execution are separate capabilities. Minirun can discover, size, download, and verify a recognized container before a model-specific tokenizer or input pipeline and bounded runner exist for that family.
- Read the support state in model details. Storage and verification only is a terminal description for this app version, not a request to verify again.
Storage and downloads
Why is a drive listed under Volumes while Storage says no location has been added?
Volumes describes mounted disks. A storage location is a folder you explicitly grant through the system picker. Detecting a disk does not grant Minirun access to every folder on it.
- Choose Add Folder in Storage.
- Select the folder containing your Minirun containers and allow the scan to complete.
Why is an external SSD slow even though it uses USB-C?
USB-C describes the connector, not the negotiated data mode. The Apple device, cable, hub or dock, enclosure bridge, and SSD form one path, and the slowest hop can reduce it to USB 2 or another fallback mode. An enclosure's advertised peak is not an end-to-end measurement.
- Check the exact Mac, iPhone, or iPad technical specifications for its USB or Thunderbolt mode.
- Use a cable explicitly rated for that data mode, bypass unnecessary hubs, and confirm the enclosure supports the same mode or an intentional fallback.
- If the path remains stable but slower than expected, treat that as an observed link result rather than changing the memory budget to hide it.
Why does an external drive disconnect or disappear during a long operation?
Common path-level causes include insufficient bus power, a charge-only or marginal cable, an unstable hub, enclosure sleep or firmware behavior, and thermal shutdown. Minirun cannot safely continue a read from a volume the operating system has removed.
- Stop the operation and restore a dry, firmly connected storage path.
- Use a short cable rated for the intended data mode. For iPhone or an unstable bus-powered enclosure, use a wall-powered USB-C hub or dock.
- Check the enclosure vendor's firmware and thermal-pad instructions. Do not use cooling that creates condensation.
- After the volume is stable, reauthorize the same folder if requested and run full verification before chat.
Can I use a Thunderbolt or USB4 NVMe enclosure with iPhone?
Only if the enclosure explicitly supports a USB 3 fallback understood by the iPhone. A USB4 or Thunderbolt bridge does not make the iPhone a Thunderbolt host; the phone uses the mode in its own technical specification. Thunderbolt-only storage is therefore not a safe purchase for this path.
- Check the enclosure controller and finished-product specification for USB 3.2 backward compatibility.
- Use a 10 Gb/s-rated USB-C data cable and external power when required.
Why did a paused download not continue after relaunch?
Download jobs are durable, but Minirun does not restart network activity without an explicit request. A restored Paused job is expected to remain stopped.
- Mount the destination storage if necessary.
- Open the saved transfer and choose Resume. If destination authority is stale, reselect the folder when prompted.
What should I do after a verification or storage-read error?
Use the named failure and verification report. Repair applies only when a complete report identifies files that can be fetched again. A missing drive, changed repository identity, unsafe path, or unsupported container needs its own correction.
- Keep the job stopped and restore access to the same storage location.
- Rescan, then run full verification. Repair only the files the report identifies.
Runs and Instruments
Why can a run spend a long time in prefill?
Prefill processes the prompt before generated tokens are available. Its duration depends on the model, prompt length, current memory plan, storage path, and device conditions. A long prefill is not by itself a stall.
- Open Instruments and check whether the current stage or layer boundary is advancing.
- Confirm that the model drive remains mounted and that the memory budget is at or above the displayed floor.
- If the stage stops advancing and a named error appears, use that error rather than the elapsed time as the diagnosis.
Why does Instruments show a dash, Waiting, or Not reported?
The runtime has not yet produced the boundaries needed for that metric, or the selected model runtime does not report it. These are unknown evidence states, not measured zeros.
- Wait until the relevant stage crosses the boundaries named by the panel.
- If the run finishes without the metric, treat it as unavailable for that run and runtime.
How do I compare two runs?
Keep the model and container revision, prompt, memory budget, storage path, device and OS, and relevant cache state constant. Compare equivalent stages and report variability across fresh runs rather than selecting one favorable result.
- Record the complete run configuration before starting.
- Change one variable at a time and repeat both conditions under comparable device and storage conditions.