Troubleshooting

A model download is slow or seems stuck

Models are large — several gigabytes for the bigger ones. The download continues while the app is in the background and reports progress as it goes.

If it has genuinely stalled, cancel and restart it: downloads resume rather than starting over.

On a metered or slow connection, start the download somewhere you can leave it running.

The app is killed during a long conversation

The conversation cache, not the model, is usually what ran out of memory. See why a long chat costs more than the download.

Switching to a smaller model, or starting a fresh conversation, both fix it. If you have documents enabled and are not using them, turning retrieval off releases the embedding model too.

The Mac's models are not in my picker

Check, in order:

1. The Mac app is running, with sharing on.
2. Both devices are on the same network — or the phone has a manual address that reaches the Mac.
3. The pairing completed. If you are not sure, re-pair: it takes one code.

You should not need to open the pairing screen for a Mac you have already paired. The phone reconnects on its own when the app opens.

A reply about my document said almost nothing

Ask something more specific. Given a vague prompt and a page of evidence, a small model sometimes produces a fragment rather than a sentence. "When does it expire?" works far better than a word or two.

If a reply comes back as nothing but a citation marker, the app replaces it with an explanation rather than showing it — a bare [1] looks like a rendering bug rather than a model that said nothing.

My document was added but answers ignore it

A router decides per message whether your documents are relevant, so a question it reads as general knowledge will not consult them. Phrasing the question around the document's subject usually settles it.

Check also that the document actually indexed: a PDF whose pages are images is recognised on the device, but a file with no readable text at all is reported rather than stored as empty.

Generation is slower than it was

Two normal causes:

• The first generation after loading a model is slower, because Metal compiles and caches its shaders on first use.
• A warm phone throttles. Sustained generation produces heat, and the system reduces clocks in response.

Neither indicates a fault.

Still stuck

Email support@elythros.com. The model you were using, your iOS version, and what you did just before it went wrong are the three things that make a problem fixable.