When an AI model won’t download to your phone, the cause is almost always one of four things: not enough free storage, a connection that dropped or went to sleep, a link that points at a web page instead of the model file, or a Hugging Face repository that requires a login. Each has a quick fix, and the error message usually tells you which one you have.
This guide covers downloads specifically. If the file arrived and the trouble starts when the model loads, that’s a memory problem instead: see local AI app crashing or model won’t load.
Quick checklist #
- Free up storage. You need the model’s size plus a comfortable margin, and more if you add image support.
- Use stable Wi-Fi and keep the phone plugged in for big models.
- Keep the app open until the download finishes, or pause it before you switch away.
- Check the link if you added a custom model: it must be the direct file link, containing
/resolve/and ending in.gguf. - Pick a model your phone can run, so the download isn’t wasted. Look for the “Fits your device” badge.
If you’re still stuck, find your symptom below. The examples use Personal LLM, which downloads models straight from Hugging Face, but the causes apply to any local AI app.
“Not enough storage” or the download won’t start #
Phone models are big files. In the Personal LLM catalog they range from 0.81 GB (Qwen 3.5 0.8B) to 6.17 GB (GLM 4.6V Flash), and adding image support to a model is another 195 MB to 1 GB. The app checks free space before a download begins and won’t start one that can’t finish.
To make room:
- Delete models you no longer use from the Models tab. You can always download them again.
- Offload or delete large apps you rarely open.
- Clear downloaded videos from streaming and podcast apps, which quietly take several gigabytes.
- On iPhone, check Settings > General > iPhone Storage; on Android, Settings > Storage.
Leave space spare after the download, too. A phone that’s completely full slows down and can fail to load the model afterwards. How much storage AI models need has sizes and a sensible budget.
One iPhone trap: “Offload App” removes the app but keeps its documents and data, so offloading a local AI app won’t free the space its models use.
The download stalls, fails halfway or restarts #
Multi-gigabyte downloads take a while, and phones are aggressive about pausing work that isn’t on screen.
Stay on Wi-Fi. A switch from Wi-Fi to mobile data, or a weak signal, can break the connection, and many phones limit large downloads on cellular by default.
Keep the app in front, or pause first. Personal LLM lets you pause, resume and cancel. If you need to leave the app, tap pause first. What happens after an interruption differs by platform:
- Android: if the app is closed or the system kills it mid-download, the partial file is kept and the model card offers Resume on the next launch.
- iPhone: a download you paused resumes where it stopped. If the app is force-quit without pausing, that file starts again from the beginning.
Turn off Low Power Mode or battery saver during the download, since those modes cut background network activity.
Try again later. Hugging Face occasionally has slow spells. A download that fails with a server error often succeeds a few minutes on.
A custom model link doesn’t work #
If you’re adding your own GGUF by URL, the link itself is the most common problem. Our full guide to running a custom GGUF model on your phone walks through picking one; here are the errors people actually hit.
You copied the page link, not the file link #
On Hugging Face, each file has two addresses:
| Link type | Looks like | Works? |
|---|---|---|
| File page (for viewing) | huggingface.co/user/repo/blob/main/model-Q4_K_M.gguf | No: this is a web page |
| Direct download | huggingface.co/user/repo/resolve/main/model-Q4_K_M.gguf | Yes |
Both end in .gguf, so the mistake is easy to make. The fix: on the file’s page, long-press the download button (the small arrow icon) and copy that link, which contains /resolve/. Personal LLM’s add-model screen also requires the link to start with https:// and end with .gguf.
“The server refused the download (HTTP 401 or 403)” #
That’s a gated repository. Some publishers require you to sign in to Hugging Face and accept a license before downloading. An app that downloads anonymously can’t get past that screen, and a well-behaved app tells you so rather than saving a broken file.
Look for another repository publishing a GGUF of the same model without a gate. Stick to well-known uploaders, and read whether AI model files are safe to download before picking one.
“The server returned HTTP 404” #
The file isn’t at that address. Check for a typo, and check that the file hasn’t been renamed or removed from the repository. Repository owners do reorganize files, and old links go stale.
The model is split into several files #
Very large GGUF models are sometimes split into parts named like -00001-of-00003.gguf. Adding one part won’t give you a working model. For a phone, choose a smaller quantization that comes as a single file, which is usually the right size anyway.
It downloaded, but something’s wrong #
Once the file is on the phone, the failure modes change. Quick triage:
- The app closes while the model loads, or on the first long reply. That’s memory, not the download. Try a smaller model, close other apps, lower the context size: local AI app crashing fix.
- “Model file not found” or it won’t open at all. The file is incomplete or was removed. Delete the model and download it again.
- It loads but writes gibberish or repeats forever. Usually the wrong prompt template or a base model instead of an instruct one, which why local AI repeats itself covers.
How to update or replace a model #
Downloaded models never update themselves. When a newer version appears, download it as a separate model, check it works on questions you care about, then delete the old one to reclaim the storage. Your chats stay in the app either way, and you can switch any conversation to the new model from the chat header. Do offline AI models get updated covers when an upgrade is worth the download.
Frequently asked questions #
Why won’t my AI model download? #
Usually low storage, an unstable connection, or a link to a web page instead of the file. For custom models, a gated Hugging Face repository that requires a login also blocks the download. Work through them in that order.
Can I download AI models over mobile data? #
You can, but models are 1 to 6 GB each, so Wi-Fi is strongly recommended. Some phones also restrict large downloads over cellular unless you change the setting.
Why does my model download start over on iPhone? #
iOS only keeps resume information when a download is paused properly. If the app is force-quit or closed by the system mid-download, that file starts again. Pause before leaving the app, and keep the phone on Wi-Fi and charging.
How do I download a gated model? #
Apps that download anonymously can’t pass the Hugging Face login and license screen. Look for a GGUF build of the same model in an ungated repository from an uploader you trust.
Does a failed download leave junk on my phone? #
A partial file may remain so the download can resume. If you’ve given up on that model, cancel or delete it from the Models list rather than leaving the fragment, and the space is returned.