To run a local LLM on Android, install an app that runs open models on the phone, check how much RAM your phone has, download a model that fits, and start chatting. Once the model is downloaded, it works in airplane mode. Budget phones with 3 to 4 GB of RAM can run small models. Phones with 8 GB or more, which includes most recent flagships, can run 9-billion-parameter models.
No root, no sideloading and no computer needed. The apps are on Google Play.
How much RAM does your Android phone have? #
RAM decides which models you can run. To find it:
- Open Settings → About phone. Many phones list RAM there, sometimes under Hardware or Device specifications.
- On Samsung phones, Settings → Device care → Memory shows total RAM.
- If you can’t find it, search your model name plus “RAM” on the manufacturer’s spec page.
Some phones advertise “virtual RAM” or “RAM Plus” that uses storage as extra memory. Ignore it for AI purposes: only physical RAM helps a model run at a usable speed.
| Phone RAM | Typical phones | What runs well |
|---|---|---|
| 3 to 4 GB | Budget and older phones | Models under 1B to about 3B |
| 6 GB | Mid-range | 3B to 4B models |
| 8 GB | Upper mid-range, older flagships | 4B comfortably, 9B at a squeeze |
| 12 GB or more | Recent flagships | 9B models, including vision models |
For the reasoning behind these numbers, see how much RAM you need to run an LLM on a phone.
Does your phone’s chip matter? #
Yes, for speed. Local AI apps use the GPU when they can, which is much faster than the CPU alone.
- Qualcomm Snapdragon phones have Adreno GPUs. Apps built on llama.cpp can use OpenCL on Adreno 700-series GPUs and newer, which covers Snapdragon flagships of the last few years.
- Google Tensor (Pixel), Samsung Exynos and MediaTek Dimensity phones generally run these apps on the CPU. That works, just more slowly, so a slightly smaller model may feel better.
Speed also depends on cooling. Phones throttle when hot, so a model that starts fast can slow down during a long answer.
Step by step: your first local model on Android #
- Install a local AI app from Google Play (options below).
- Connect to Wi-Fi. Models are 1 to 6 GB.
- Choose a model for your RAM. A 3B or 4B model is the right start for most phones.
- Download it. Keep the app open or let it resume if the connection drops.
- Switch on airplane mode and ask something. If it answers, it’s running on your phone.
- Check the speed. If replies are slow, try a smaller model; if they’re fast, try the next size up.
Which model should you start with? #
- 3 to 4 GB: Gemma 4 E2B (about 2 GB) or Ministral 3 3B. Qwen 3.5 0.8B is the fallback for very old phones.
- 6 GB: Qwen 3.5 4B, the best balance of speed and quality at this size.
- 8 GB: Gemma 4 E4B or Qwen 3.5 4B for speed; Qwen 3.5 9B if you’ll trade speed for better answers.
- 12 GB and up: Qwen 3.5 9B for general use, GLM 4.6V Flash for documents, screenshots and photos.
Our guide to the best open-source LLMs for phones compares them in more detail.
Apps that run local LLMs on Android #
| App | Price | Notes |
|---|---|---|
| Personal LLM | Free, ad-supported | Curated catalog with device-fit badges, vision, document chat, custom GGUF by URL |
| PocketPal AI | Free, open source | Browse and download GGUF models from Hugging Face |
| Google AI Edge Gallery | Free | Google’s app for trying Gemma models on-device; Android 12 or later |
Personal LLM runs on Android 7 or later. It reads your phone’s actual RAM and labels every model before you download it: “Fits your device” when you have at least 2 GB of headroom, “Should run” when you meet the minimum. It uses llama.cpp with OpenCL on Snapdragon Adreno 700-series GPUs and newer, and quietly falls back to the CPU on other chips. A tokens-per-second readout on each reply and a built-in benchmark let you see exactly how your phone performs. Downloads pause and resume, and the app checks free space before starting.
What about Gemini Nano? #
Some Pixel and Samsung phones include Gemini Nano, a small Google model that runs on the device for specific features such as summaries and smart replies. It’s a useful built-in, but you don’t choose the model, you can’t hold an open-ended conversation with it the way you can with a chat app, and it’s only on certain phones. A local LLM app works on a much wider range of Android phones and lets you pick the model. For what your phone’s built-in AI does with your data, see phone AI privacy settings.
Tips for speed and battery on Android #
- Exclude the app from battery optimization if downloads stop when the screen turns off. Settings → Apps → the app → Battery.
- Close memory-heavy apps such as games before a long session.
- Turn off battery saver when you want full speed.
- Use thinking mode only when you need it. It improves answers on math and logic but takes longer.
- Keep the phone cool. Take it out of a thick case or off a charger if it’s throttling.
- Delete unused models to recover storage.
Frequently asked questions #
Can Android run an LLM offline? #
Yes. After you download a model in a local AI app, it runs entirely on your phone’s processor and doesn’t need a connection. Only the model download requires internet.
How much RAM do I need to run an LLM on Android? #
About 3 GB for the smallest useful models, 6 GB for good 4B models, and 8 GB or more for 9B models. Physical RAM is what counts; “virtual RAM” features don’t help.
Why is my local AI slow on Android? #
Usually the model is too big for your phone’s RAM or chip, the phone is running on the CPU rather than the GPU, or it’s hot and throttling. Try a smaller model, close other apps and let the phone cool down.
Do I need to root my phone to run a local LLM? #
No. Apps on Google Play run open models without root or special permissions. You just need enough RAM and free storage for the model.