How Much Storage Do AI Models Need on Your Phone?

A language model small enough to run on a phone takes between about 0.8 GB and 6 GB of storage. A good everyday model of around 4 billion parameters needs roughly 2.5 to 3 GB. Adding image understanding costs another 195 MB to 1 GB. Built-in AI takes space too: Apple says Apple Intelligence needs 7 GB of free storage.

Storage is the easy half of the question. The model also has to fit in your phone’s RAM while it runs, and that’s usually the tighter limit, which how much RAM you need to run an LLM on a phone covers in full.

How big are phone AI models? #

These are the models in the Personal LLM catalog with their download sizes, typical of what any local AI app offers.

Image support is a separate file per model, the vision projector, adding 195 MB to 1 GB depending on the model.

A rule of thumb for mid-size and larger models: at the compressed formats mobile apps use, each billion parameters takes a bit over 0.6 GB. Qwen 3.5 4B works out to about 0.68 GB per billion and Qwen 3.5 9B to about 0.63. Very small models run higher per billion, because parts of a model don’t shrink in proportion.

Why are the files so much smaller than the original models? #

Open models are published at full precision, where every parameter takes 16 bits. A 4-billion-parameter model at that precision is around 8 GB. Mobile apps use quantized versions in the GGUF format, which store each number with fewer bits, cutting the size by roughly three quarters for a modest loss in quality.

That’s what labels like Q4_K_M or Q8_0 mean on Hugging Face: the number is roughly the bits per weight, and lower means smaller, faster and slightly worse. For phones, 4-bit builds are the usual sweet spot. GGUF quantization explained has the sizes for each level.

How much free space should you keep? #

Plan for the model plus a comfortable buffer. Phones slow down and misbehave when storage is nearly full, and a download fails outright if the space runs out partway.

A sensible setup for most people:

1. One everyday model at 2 to 3 GB.
2. Image support for that model if you’ll ask about photos, up to 1 GB.
3. Optionally a bigger model for hard questions, 5.7 to 6.2 GB, if your phone has 8 GB of RAM or more.
4. At least 5 GB free beyond that for the system, updates and photos.

So budget about 4 GB for a light setup and 10 to 12 GB with a large model added. Personal LLM checks free space before a download starts, so it won’t leave you with a half-finished file. If a download does fail, AI model download not working has the fixes.

How to check and free up space #

iPhone: Settings > General > iPhone Storage shows what’s using space, app by app. Each entry includes its “Documents & Data,” which is where downloaded models live.

Android: Settings > Storage shows totals by category. Tap Apps to see which take the most, including their data.

Quick ways to make room:

• Delete apps you haven’t opened in months.
• Clear old video downloads from streaming apps.
• Turn on iCloud Photos with “Optimize iPhone Storage,” or back your photos up elsewhere.
• Remove offline map regions and podcast episodes you no longer need.
• Delete AI models you stopped using. They re-download later in minutes on Wi-Fi.

A warning about iPhone’s “Offload App” option: it removes the app but keeps its documents and data, so offloading a local AI app won’t free the space its models use. Deleting the app will, and in Personal LLM’s case that removes your chats too, so export a backup first if you want to keep them.

Does storage speed matter? #

Only for loading. A model is read from storage into memory once, when you first use it in a session, and after that it runs from RAM. On a phone with fast internal storage that load takes a few seconds; from slow external storage it can take much longer, which is one reason most apps keep models in internal app storage.

Is a bigger model worth the space? #

Sometimes. The jump from a sub-1B model to a 4B one is large: the tiny model is fast, but its answers are noticeably weaker. The jump from 4B to 9B is smaller but real, especially for math, code and questions you want to check carefully. A 9B model also generates more slowly, so you trade speed for quality. Local LLM speed on your phone covers that trade-off with real numbers.

For most people, one good 4B model is the right amount of storage to spend. You can add a larger one later, and Personal LLM lets you switch models mid-conversation from the chat header when a question needs more.

If your phone is short on both RAM and storage, the smallest models still work; see running AI on an old phone. And if you’re comparing apps before committing the space, our best offline AI apps roundup covers the options.

Frequently asked questions #

How many GB is a local LLM? #

Phone-sized models range from under 1 GB to about 6 GB. A 4-billion-parameter model, the usual starting point, is around 2.5 to 3 GB. Desktop models can be tens of gigabytes, but those won’t run on a phone.

Does Apple Intelligence take up storage? #

Yes. Apple says Apple Intelligence requires 7 GB of storage on the device, which holds Apple’s own on-device models. That’s separate from any models you download in a local AI app.

Can I store AI models on an SD card? #

It depends on the app and the phone. Most local AI apps keep models in their own app storage on internal memory. Even where external storage is possible, loading a multi-gigabyte model from a slow card is much slower than from internal storage.

Why does the app need more RAM than the model’s file size? #

The model needs working memory beyond its weights, mainly for the conversation’s context, and the operating system and other apps need their share. That’s why apps recommend headroom, such as the 2 GB margin behind Personal LLM’s “Fits your device” badge.

Do models get bigger over time as I chat? #

No. The model file stays the same size. Your chat history takes a few kilobytes per conversation, and photos you attach add their own size.