When an on-device AI app closes while loading a model or in the middle of a reply, the cause is almost always memory. The model, the conversation and any photo you attached have to fit in the RAM your phone allows one app, and when they don’t, iOS or Android shuts the app down without an error message. The fix is usually a smaller model, a shorter context, or closing other apps, and it takes a minute or two.
This is about what happens once the model file is on your phone. If the download itself is failing, start with AI model download not working.
Why do local AI apps crash? #
A language model runs from memory, not from storage. Roughly, the RAM you need is:
- The model itself: close to its file size. A 2.7 GB model needs about 2.7 GB.
- The conversation: the context window. Longer chats and larger context settings use more.
- Image support: vision models load a separate projector file of about 195 MB to 1 GB.
- Everything else: the operating system and every other open app.
A phone with 6 GB of RAM doesn’t give 6 GB to one app. The system keeps a large share for itself and for background apps, and when an app asks for more than is available, the system closes it. That’s why the same model loads fine after a restart and crashes after an afternoon of games and camera use.
What does your symptom mean? #
| What happens | Most likely cause | First thing to try |
|---|---|---|
| App closes while the model is loading | Not enough free RAM for the model | Close other apps, or pick a smaller model |
| Model loads, then the app closes on the first long reply | Context too large for the remaining memory | Lower the context size, start a new chat |
| App closes when you attach a photo | Vision projector plus model is too much | Use a smaller vision model, close other apps |
| First photo question takes a very long time | The phone is encoding the image on the CPU | Wait; it’s slow, not stuck |
| “Model file not found” or the model won’t open | The download is incomplete or the file was removed | Delete and re-download the model |
| Replies are extremely slow | No GPU acceleration on this phone, or heat | Smaller model, let the phone cool |
| A custom model loads but writes gibberish | Wrong prompt template, or a base model | Use a standard instruct build from a known source |
9 fixes, in the order to try them #
- Close other apps and restart the phone. This frees more memory than anything else, and it’s the fastest fix when a model that used to work suddenly doesn’t.
- Pick a model with headroom. In Personal LLM, each model card shows a badge read from your phone’s actual RAM. “Fits your device” means at least 2 GB of headroom over the model’s minimum; “Should run” means it only just meets it. “Should run” models are the ones that crash when other apps are open, so drop a size (from Qwen 3.5 9B to Qwen 3.5 4B, say) if crashes keep happening.
- Lower the context size. The context setting controls how much of the conversation the model holds at once. Personal LLM defaults to 2,048 tokens and goes up to 8,192; higher values cost memory, so bring it back down if crashes started after you raised it. Vision chats are capped at 4,096 for the same reason.
- Start a new chat. A long conversation fills the context. Starting fresh for a new topic is lighter on memory, and the answers are more focused anyway.
- Use a smaller model for photo questions. Image support adds a second file to memory. On a phone with 4 to 6 GB of RAM, a 4B model with image support is far more reliable than a 9B vision model.
- Be patient with the first image on older Android phones. GPU acceleration uses OpenCL on Snapdragon phones with Adreno 700-series or newer graphics and falls back to the CPU elsewhere, including on Exynos Galaxy phones. On the CPU, the first photo takes a while to process. Give it time before assuming it froze.
- Re-download a model that won’t open. An interrupted download or a full disk can leave a broken file. Delete the model and fetch it again.
- Let a hot phone cool down. Heat doesn’t usually cause crashes, but a throttled phone gets so slow it looks hung. See why your phone gets hot running AI.
- Update the app. Inference engines improve memory handling and add support for new model types with each release.
Problems with custom GGUF models #
If you added your own model file, a few extra things can go wrong:
- The file is too big. A 12B or 14B model at Q4 needs more RAM than most phones will give one app. Stay at about 9B and below unless you have 12 GB or more.
- The quantization is wrong for a phone. Q8_0 and F16 files are much larger than Q4_K_M for a modest gain. GGUF quantization explained covers which to pick.
- The architecture isn’t supported yet. Brand-new model families sometimes need an engine update before any app can run them.
- It’s a base model, not an instruct model. Base models continue text rather than answering, which reads as broken output.
- The vision projector doesn’t match. A projector only works with the exact model it was built for.
The safe way to add one is in run a custom GGUF model on your phone.
Before you reinstall, back up your chats #
Reinstalling fixes a genuinely broken setup, and it also deletes everything, because an on-device app keeps your chats only on the phone. In Personal LLM, export a backup first: every chat and setting goes into one JSON file you can keep and restore later from the file or the clipboard. Models have to be downloaded again afterwards, so do it on Wi-Fi.
How much memory does your phone have? #
On iPhone, Apple doesn’t show RAM in Settings, but it’s roughly 4 GB for the iPhone 13, 6 GB for the iPhone 14 and 15, and 8 GB or more from the iPhone 15 Pro and iPhone 16 onward. On Android, look under Settings > About phone, or in the device specs. Apps that show a fit badge read the real number for you. For what each amount can run, see how much RAM you need to run an LLM on a phone.
Frequently asked questions #
Why does my AI app crash when I send a photo? #
Answering about a photo loads an extra image-projector file on top of the model, and the photo itself takes memory to process. If the phone was already near its limit, the system closes the app. Use a smaller vision model, close other apps, and keep the context size at the default.
Why does the model load but crash on the first message? #
The model fit, but the model plus the conversation’s working memory didn’t. Lowering the context size, starting a new chat, or choosing a smaller model usually fixes it.
Is a crashing AI app a sign of malware? #
No. On-device AI apps are among the most memory-hungry apps you can run, and the operating system closes them when memory runs out. That’s normal behavior, not a sign of anything wrong with your phone.
Will a phone with more RAM stop the crashes? #
For crashes caused by memory, which is most of them, yes. More RAM lets you run bigger models and longer conversations. A phone with 8 GB comfortably runs 4B models with image support and handles 9B text models.
Why is the first reply after loading so slow? #
The phone reads the model from storage into memory and sets it up before the first word. Vision models take longer because they also load the image projector. Later replies in the same session are much faster.