A recent phone, tablet or other mobile device can run a small language model on the device itself. Chats stay on the device, and they work with no connection once the model is downloaded. Getting a model onto a mobile device is different from a computer: there are no commands to paste. You install a chat app, then download one model file through it.
1. Which devices this covers
Anything that runs iOS, iPadOS or Android and has enough RAM. The brand and the shape matter much less than the memory and the chip.
- Phones and foldables: iPhones and Android phones from any maker, including fold and flip models.
- Tablets: iPads and Android tablets. They use the same apps and the same steps as phones, and many have more RAM.
- Android handheld game consoles: they run ordinary Android apps, and several have 12 GB or 16 GB of RAM and a fan.
- Chromebooks that run Android apps: if yours has the Google Play Store, the Android app installs from there. PocketPal is built for Intel and AMD chips as well as Arm ones.
- Devices without Google Play: Amazon Fire tablets and some Huawei and other devices. You install the app’s APK file instead (step 3). Most Fire tablets have 4 GB of RAM or less, so only the smallest models fit.
- Devices that mostly can’t: TV boxes and sticks, watches and e-readers. They have 1 to 4 GB of RAM and often a 32-bit system. PocketPal needs Android 7.0 or later on a 64-bit chip.
We checked the steps below on a real phone. We haven’t tested a handheld console, a Chromebook or a Fire tablet ourselves, so treat those as “should work” until you have tried it.
2. What your device can run
RAM decides it. iOS and Android let one app hold only part of the device’s RAM before they close it, and the model has to fit in that part along with its working memory (memory types). This is what we plan around:
| RAM | One app can use | Largest model, about |
|---|---|---|
| 3 GB | 1.4 GB (45%) | 0.5B parameters |
| 4 GB | 1.8 GB (45%) | 1.5B parameters |
| 6 GB | 3.0 GB (50%) | 3B parameters |
| 8 GB | 4.4 GB (55%) | 5B parameters |
| 12 GB | 7.2 GB (60%) | 10B parameters |
| 16 GB | 9.9 GB (62%) | 15B parameters |
| 24 GB | 15.6 GB (65%) | 24B parameters |
The last column is for a Q4_K_M file with a normal chat length (quantization). Storage only keeps the file: about 2.4 GB for a 4B model and 4.9 GB for an 8B model. A model too big for any mobile device says so on its page instead of showing these steps.
To find your RAM: on Android, open Settings and look under About phone, About tablet or About device. iPhones and iPads don’t show it in Settings, so find your model on our list of mobile devices, or press Detect on My hardware.
3. Install a chat app
Which app, and where it comes from, depends on how your device gets apps.
- App Store (iPhone, iPad): install PocketPal AI. It is free.
- Google Play (most Android phones, tablets, handhelds and Chromebooks): install PocketPal AI. It is free and is our first pick on every device.
- No Google Play: download the APK file from PocketPal’s releases page on GitHub and open it. Android asks you to allow installs from your browser first. ChatterUI is a second Android app that installs the same way, from its own releases page; it isn’t on Google Play at all.
Draw Things makes images on an iPhone or iPad (image generation). No Android app in our list does that yet. See all of them side by side under mobile apps.
4. Get the model in PocketPal
There are two ways in. Each model page shows the one that applies and names the exact file.
Search inside the app (most chat models):
- Open PocketPal and tap the menu (☰), then Models.
- Tap +, then Add from Hugging Face.
- Search for the repository named on the model page, open it and download the file the page names.
- Tap Load next to the model and start a chat.
Download in the browser, then add the file (models that PocketPal’s search doesn’t list, and models with a license to accept):
- Tap the download link on the model page. The file goes to Files › Downloads on an iPhone or iPad, or the Downloads folder on Android.
- In PocketPal, tap the menu (☰), then Models, then + and Add Local Model, and pick the file.
- Tap Load next to the model and start a chat.
A few models are gated: sign in at huggingface.co in the device’s browser and accept the model’s license before the download link works (licenses).
5. Or use ChatterUI on Android
- Download the model file in the device’s browser. It goes to your Downloads folder.
- In ChatterUI, open the settings menu, set App Mode to Local and tap Models.
- Tap the add-file button, then Copy Model Into ChatterUI, and pick the file.
- Tap the model to load it and start a chat.
We checked the PocketPal steps on a real phone. We haven’t yet done the same for ChatterUI, so a menu name may differ in your version.
6. Use one whole file
Mobile apps open a model as a single .gguf file (formats). Larger files on Hugging Face are often split into parts, with names ending in something like -00001-of-00003.gguf, and mobile apps can’t open those. So our steps name one whole file: Q4_K_M when it comes as one file, or else the largest whole file that fits your device. When a model has no whole file that fits, its page says no mobile app runs it.
If your device isn’t in our list
We list well-known phones and tablets by name. For anything else, the chip and the RAM are what count:
- You know the chip: pick the entry named “Other Android device with” that chip (Snapdragon, Dimensity or Exynos) under Mobile devices, then choose your RAM.
- You don’t: pick “Other Android device”, “Other iPhone” or “Other iPad” and choose your RAM. The memory answer is right; the speed is a mid-range estimate.
- Or skip the list: the memory calculator takes a model and an amount of memory.
What to expect
- Speed follows the device’s memory speed (why). Each device’s page gives an estimate for every model.
- Long prompts start slowly on Android. The apps read the prompt on the CPU there, so pasting in a long document takes much longer to answer than on an iPhone or iPad (two speeds).
- If the app closes while loading or mid-chat, the model is too big for what the system lets one app use. Close other apps, lower the context length in the app’s settings (why), or pick a smaller file or model.
- Heat and battery. Generating text keeps the chip busy. A phone gets warm and slows itself down in long sessions, and the battery drains faster than usual. Tablets and handhelds with a bigger body, a bigger battery or a fan hold their speed longer.
- Offline means offline. After the download you can switch on airplane mode and keep chatting (what local guarantees).
Using this site on a mobile device
Open a model page on your phone or tablet and it shows the steps for your system: the store button, the taps and the file to get. The commands for a computer are still there, behind Show commands for a computer. Pages such as “Can my device run this model?” do the same for one device and name the file that suits its memory.