A recent phone has enough processing power to run a small AI model by itself. PocketPal is a free app that does exactly that: you download a model once, and after that it answers questions on the phone, with Wi-Fi and mobile data switched off.
PocketPal, Gemma and Qwen are not made by webridgeAI. What the app does was checked against its GitHub page and its App Store listing, and the model sizes against the Hugging Face pages linked in section 4, on 17 October 2026. Apps and model pages change, so check the sizes again before a large download.
1. What PocketPal is
| What | Detail |
|---|---|
| The app | PocketPal AI. Free to install and use. The App Store lists optional in-app purchases for extra assistants called Pals; you do not need them to download a model and chat. |
| Licence | MIT. The source code is public on GitHub, so anyone can read it. |
| Phones | iPhone and iPad (the App Store listing says iOS 15.1 or later) and Android. |
| How it runs | On the phone itself, using its processor and, where the phone supports it, its graphics chip. Nothing is sent to a server to produce an answer. |
| Models | It runs models in the GGUF format, and downloads them from Hugging Face. The project names Gemma, Qwen, Phi and Llama. |
| Account | None. The project's README says "No account, no cloud, no internet required." |
2. Install it
Use the links from the project's GitHub page, because other apps have similar names:
- iPhone or iPad: PocketPal AI on the App Store. The app itself is about 88MB.
- Android: PocketPal AI on Google Play.
On first launch the app shows a short introduction. In the current version it then offers a recommended model for your phone, with your phone's RAM and free storage shown on the screen. You can take that suggestion, or follow the table in section 4. The App Store build may lag the GitHub release, so menus and screens can differ slightly from what this guide describes.
3. Find out how much memory your phone has
A model has two sizes that matter. Storage is where the file lives while it is downloaded. RAM is the workspace the phone needs while the model is answering, and it is the number that decides which model fits.
The easiest way, on iPhone and Android, is to ask PocketPal. The recommended-model screen in its first-run setup shows your phone's RAM, and its Benchmark page (open the menu and look for Benchmark) lists the total memory of your device.
Or look it up yourself:
- iPhone: Settings does not show RAM. Open Settings, General, About, note your Model Name, and search for that model with the word RAM.
- Android: Open Settings, then About phone. Many phones list RAM there. If yours does not, search for your phone's model name with the word RAM. If the phone advertises "virtual RAM", "RAM Plus" or "extended RAM", ignore that figure and use the physical one.
If your number falls between two rows of the table below, take the smaller row.
4. Choose the model for your phone
Pick one model from your row. Each step down the table is a smaller file, which needs less memory and runs faster but answers less well. Gemma is Google's family of open models and Qwen is Alibaba's. If you are unsure, take the smaller file.
| Your phone | Model | Parameters | File to download | Size |
|---|---|---|---|---|
| About 4GB of RAM (3 to 5GB) | Gemma 3 1B | 1.0B | gemma-3-1b-it-Q4_K_M.gguf | 0.81GB |
| About 4GB of RAM (3 to 5GB) | Qwen3 0.6B | 0.6B | Qwen_Qwen3-0.6B-Q4_K_M.gguf | 0.48GB |
| About 6GB of RAM (5 to 7GB) | Gemma 3 4B | 4B | gemma-3-4b-it-Q4_K_M.gguf | 2.49GB |
| About 6GB of RAM (5 to 7GB) | Qwen3 1.7B | 1.7B | Qwen_Qwen3-1.7B-Q4_K_M.gguf | 1.28GB |
| 8GB of RAM or more | Gemma 4 E2B | 2.3B effective (5.1B with embeddings) | gemma-4-E2B-it-Q4_K_M.gguf | 3.11GB |
| 8GB of RAM or more | Qwen3.5 4B | 4B | Qwen3.5-4B-Q4_K_M.gguf | 2.74GB |
| 12GB of RAM or more (optional) | Gemma 4 E4B | 4.5B effective (8B with embeddings) | gemma-4-E4B-it-Q4_K_M.gguf | 4.98GB |
Parameters and sizes are from each model's Hugging Face page on 17 October 2026. What the columns mean:
- Parameters are the model's adjustable numbers, a rough measure of how much it can hold. More parameters usually means better answers and a bigger file. The "E" in Gemma 4's E2B and E4B stands for effective: 2.3B parameters do the computing, and the file also carries extra embedding weights, which is why it is larger than the name suggests.
- Q4_K_M is a compressed version of the model, stored with about four bits per number. It is the usual balance of size and quality. Q6 and Q8 files are larger and a little more accurate; Q2 and Q3 files are smaller and noticeably worse.
- Licence. Gemma 4, Qwen3 and Qwen3.5 are released under Apache 2.0. Gemma 3 is released under Google's Gemma terms; read the licence on its page if you plan to use it for work.
Gemma files on Hugging Face can be gated. Google's own Gemma 3 pages (the ones under google/ on Hugging Face) are gated: you sign in with a free Hugging Face account and accept Google's terms on the model page before you can download anything. When PocketPal needs that, it asks for a Hugging Face access token, which you create under Settings, Access Tokens on huggingface.co and paste into PocketPal's settings. The Gemma links in the table go to community copies, which were not gated when we checked on 17 October 2026, but if a search lands you on a google/ page, do the sign-in first. Qwen files are not gated, so they download with no sign-in.
Two cautions:
- The file size is not the memory it needs. The model is loaded into RAM along with working space for the chat. PocketPal's own device rules list about 1.8GB to 2GB of memory for Gemma 3 1B and about 4.1GB to 4.5GB for Gemma 3 4B, well over twice the file size. That is why a 2.5GB file does not belong on a 4GB phone.
- The chip matters as well as the RAM. The same device rules also look at your phone's processor, and place a 6GB phone with an older or budget chip one level lower. If your phone is a few years old, drop one row.
On Android, PocketPal's built-in list prefers Q4_0 files for the larger models (for example Gemma 3 4B at 2.37GB). A Q4_0 file of the same model is slightly smaller and is a sensible choice there.
5. Download and load the model
- Connect to Wi-Fi, and check you have free storage for the file.
- Open the menu (the three-line icon) and go to Models.
- If the model you want is in the list, tap Download. If not, tap + and choose Add from Hugging Face, search for the model's name from the table (for example
gemma-3-1b-it-GGUF), open the repository and pick the file ending inQ4_K_M.gguf. - Keep PocketPal open until the progress bar finishes.
- Tap Load, or tap the chevron icon on the chat screen. Loading takes a few seconds, longer for bigger models.
- Go to the chat screen and say hello. Under each reply, PocketPal shows the speed in tokens per second, so you can see how your phone does.
If PocketPal shows a memory warning when you load a model, it is telling you the model may be too big for your phone. Go back and choose the smaller file.
6. The airplane-mode test
This is the part that proves it works with no signal.
- With the model loaded, ask one question and read the answer. This is your baseline.
- Switch on airplane mode. On an iPhone, open Control Centre and tap the airplane icon. On Android, pull down the quick settings and tap Airplane mode.
- Check that the Wi-Fi and mobile data icons are off. On some phones Wi-Fi can be switched back on while airplane mode is on.
- Ask the same question again. The answer should stream in, word by word, exactly as before.
For a stricter version, keep airplane mode on, close PocketPal completely, open it again, load the model and ask a new question. Everything you need is already on the phone.
7. Three prompts to try
A small model does well with short, clear jobs. These three are good first tests.
1. Explain something plainly:
Explain how a refrigerator works in four short sentences. Use plain words and no jargon.2. Fix a message:
Rewrite this message so it is polite and clear. Keep it to two sentences.
Message: [paste your message]3. Make a checklist:
Make a checklist of 10 things to pack for a three-day trip in cold weather.
One short line per item.Each one asks for a clear format and a short answer, which is where small models are strongest. Ask them for facts you cannot check (a date, a price, a quote) and you are more likely to get a confident wrong answer.
8. When something goes wrong
| What you see | What to try |
|---|---|
| A memory warning when you load | Choose the next smaller file in the table. |
| The app closes while loading or answering | The model is too big for your phone. Drop one row, close other apps, and restart the phone. |
| Replies are very slow | Use a smaller model. Start a new chat, because long chats are slower. Keep the phone cool and plugged in. |
| A Qwen model writes a long "thinking" section before it answers | If the chat screen has a thinking switch, turn it off. Thinking makes replies much longer on a phone. |
| The download stalls | Stay on Wi-Fi, keep PocketPal open and check free storage, then start it again. |
| You are short of storage | Open Models and delete the models you no longer use. Each one takes up its full file size. |
| It forgets the start of a long chat | Start a new chat. A model can only keep a limited amount of text in mind at once. |
9. Privacy: what stays on your phone
PocketPal's README says: "No account, no cloud, no internet required." It also says the only data that leaves the phone is what you choose to share, such as benchmark results for its public leaderboard or feedback you send.
Two things to keep in mind:
- The airplane-mode test shows that chatting works with no connection. It does not prove what any app does when it is online. We have not audited the code, so treat the README's description as the project's own statement.
- Some optional PocketPal features, such as web search, need a connection by their nature. The offline claim is for plain chat with a model you have downloaded.
Honest notes
- It is a smaller model than ChatGPT. The models in this guide run from 0.6 billion to about 5 billion parameters. The models behind online chatbots are much larger, and their makers do not publish exact sizes. Expect plainer answers and weaker reasoning.
- It is slower. Replies stream in more slowly than online, and the gap grows with the size of the model and the age of the phone.
- It can be wrong. It can invent facts, names and numbers in a confident voice. It has no live information, so it does not know today's news. Check anything that matters.
- Older phones struggle. Running a model works the processor hard, so the phone may warm up and the battery will drain faster than in normal use.
- Our table is a starting point, not a guarantee. Download size is not the memory a model needs, and phones with the same RAM differ in speed. If a model is too heavy, go down one row.