To run an AI language model on your phone, you need roughly twice the model’s file size in total RAM, and a little more for larger models. In practice that means about 3 GB of RAM for a small model around 1.3 GB, about 6 GB for a model around 2.7 GB, and 8 GB or more for models in the 5 GB range. Storage matters too, but RAM is what decides whether a model runs at all.
Why RAM matters more than storage #
A language model is a large file of numbers called weights. To generate a reply, the phone has to load those weights into memory and read through them for every word it writes. Storage only holds the file. If there isn’t enough free RAM to load it, the app crashes or the operating system closes it.
On top of the model itself, the phone needs RAM for:
- The conversation context. The model keeps a working cache for the text it’s currently reading, which grows with the context length.
- The app. Its interface, images and database.
- The operating system and other apps. Your phone never gives one app all of its memory. iOS also limits how much memory a single app may use, and apps have to request a higher limit for large models.
That overhead is why the rule of thumb is about double the file size, not the file size plus a bit.
How big are AI models that run on phones? #
Model size depends on two things: the number of parameters (the weights) and how precisely each one is stored. Phone apps almost always use quantized models, which store each weight in fewer bits. The open-source llama.cpp engine, used by many on-device apps, supports integer quantization from 1.5-bit to 8-bit “for faster inference and reduced memory use”.
At 4-bit quantization, the most common choice for phones, a model takes roughly 0.6 GB per billion parameters. So:
| Model file size | Typical parameter count (4-bit) | RAM needed (rough) | Which phones |
|---|---|---|---|
| About 1-1.5 GB | 1-2 billion | About 3-4 GB | Most phones from the last several years |
| About 2.5-3 GB | 3-4 billion | About 6 GB | Many mid-range and high-end phones |
| About 4.5-5.5 GB | 7-9 billion | About 8 GB or more | High-end phones |
| 10 GB and up | 14 billion or more | 12-16 GB or more | Few phones; a computer is more practical |
These are estimates. Actual needs vary with the app, the context length and the quantization method.
How to check how much RAM your phone has #
- Android: Most phones show RAM under Settings, then About phone (the exact path varies by manufacturer). The phone’s spec sheet on the manufacturer’s site lists it too.
- iPhone: Apple generally doesn’t list RAM in its tech specs, so look up your model in a reliable spec database, or use an app that reads it. Some on-device AI apps check it for you and only offer models your phone can handle.
Total RAM isn’t the same as free RAM. A phone with 6 GB might have only 3 or 4 GB available to one app once the system takes its share, which is why the table above is conservative.
What else affects on-device AI performance #
- The chip. Newer processors, GPUs and neural engines generate text faster.
- Memory bandwidth. Reading the whole model for every word is limited by how fast memory moves data, so larger models are slower even when they fit.
- Heat and battery. Generating text is hard work for a phone. Long sessions warm it up and use battery faster than ordinary apps.
- Storage space. You need room for the model file, plus some headroom for updates.
How Xin matches models to your phone #
Xin, an on-device AI companion app, reads your phone’s memory and only offers the models it can actually run:
| Xin model | Download size | RAM needed | Plan |
|---|---|---|---|
| Xin Core | About 1.3 GB | About 3 GB | Free |
| Xin Plus | About 2.7 GB | About 6 GB | VIP |
| Xin Max | About 5.3 GB | About 8 GB | VIP |
Models your phone can’t fit are greyed out. The download happens once, doesn’t start on its own over cellular data, and after that every conversation runs offline. The larger models are sharper and remember more accurately, but they’re slower and only help if your phone has the memory for them. Xin Core is the one to start with on most phones.
For the broader case for running AI locally, see our offline AI chatbot explainer and the privacy side in private AI chat apps.
Frequently asked questions #
Can any phone run an AI model locally? #
Most phones from the last several years can run a small model of 1-2 billion parameters, as long as they have around 3 GB of RAM and a few gigabytes of free storage. Older or very low-end phones may be too slow or run out of memory. Larger models need high-end phones.
Does running AI on your phone drain the battery? #
It uses more power than most apps while it’s generating text, and long sessions can warm the phone. Short chats have a modest impact. Plugging in during long sessions helps.
Is on-device AI as good as cloud AI like ChatGPT? #
Not for every task. Cloud services run far larger models than any phone can hold, so they’re stronger at complex reasoning and broad knowledge. For conversation, roleplay and companionship, small models do surprisingly well, and they offer privacy and offline use that cloud AI can’t. We explain how these models create characters in how AI characters work.
How much storage do I need for an offline AI app? #
Plan for the model file plus about a gigabyte of headroom. A small model needs roughly 1.5-2 GB free, while larger ones need 3-6 GB or more. The app’s model screen or store listing usually states the download size.