You can run Google’s Gemma 4 on a phone by installing an app that loads open models, downloading Gemma 4 E2B or E4B once, and chatting in airplane mode. E2B is about a 2 GB download and runs on phones with 3 GB of RAM or more. E4B is about 3 GB and wants 5 GB or more. Nothing you type is sent to Google, because the model runs on your phone’s own chip.
Here’s how to pick the right size, set it up on iPhone or Android, and what to expect once it’s running.
What is Gemma 4? #
Gemma 4 is Google DeepMind’s family of open-weight models, meaning anyone can download the files and run them on their own hardware. According to Google’s Gemma documentation, the family spans small “effective parameter” models built for phones, edge devices and browsers, up through mid-size multimodal models and larger dense and mixture-of-experts models meant for computers and servers.
The small models have a 128K-token context window. All Gemma 4 models accept images, and the smaller ones also accept audio natively. The model card says Gemma 4 was pre-trained on more than 140 languages with out-of-the-box support for 35+, and the weights are released under the Apache 2.0 license.
Only E2B and E4B are realistic on a phone. The “E” means effective: part of their parameters can be stored more cheaply, so they behave like 2B and 4B models while having more parameters in total. What 4B, 9B and E2B mean explains the naming.
Which Gemma 4 model fits your phone? #
| Model | Download | Needs RAM | Best for |
|---|---|---|---|
| Gemma 4 E2B | 2.04 GB | 3 GB+ | Older and mid-range phones, fast replies |
| Gemma 4 E4B | 3.00 GB | 5 GB+ | Better answers, harder questions |
Sizes are from the Personal LLM catalog. As a rough guide to phone RAM: iPhone 13 has about 4 GB, iPhone 14 and 15 about 6 GB, and iPhone 15 Pro and later 8 GB or more. Recent Android flagships have 8 to 12 GB. The app reads your actual RAM and marks each model “Fits your device” when there’s at least 2 GB of headroom. How much RAM you need to run an LLM on a phone has the full picture.
The larger Gemma 4 models are a different proposition: they need far more memory than a phone has, or in the mid-size case only work on a 12 GB phone loaded as your own GGUF, and slowly.
How to run Gemma 4 on iPhone or Android #
- Install Personal LLM from the App Store or Google Play. It’s free, needs no account, and runs on iOS 15.1 or later and Android 7 or later.
- Open the Models tab and find Gemma 4 E2B or E4B. Check the fit badge.
- Download on Wi-Fi. The file comes straight from Hugging Face. Downloads pause and resume if the connection drops, and the app checks free space first.
- Start a chat and pick Gemma 4 in the chat header.
- Turn on airplane mode if you want proof it’s running on the phone. It keeps answering.
- Optional: add image support, a separate download of 195 MB to 1 GB, so you can ask Gemma about photos.
Answers stream with a live tokens-per-second readout, and there’s a built-in benchmark if you want to compare E2B and E4B on your own hardware. On iPhone the app uses the GPU through Metal; on Android it uses OpenCL on Snapdragon phones with Adreno 700-series or newer graphics, and the CPU elsewhere, including on Exynos Galaxy phones.
To load a specific Gemma 4 build yourself, such as a different quantization, Personal LLM also accepts any GGUF by URL: see run a custom GGUF model on your phone.
Asking Gemma 4 about photos #
With image support installed, attach a photo or take one and ask things like:
- “What does this sign say in English?”
- “List the ingredients on this label.”
- “Which of these menu items are vegetarian?”
- “What’s wrong with this error message on my screen?”
The photo is processed on the phone and never uploaded. Small vision models can misread dense or stylized text, so double-check anything important. Gemma 4’s small models can also take audio in software that supports it, but Personal LLM uses them for text and images only; it has no voice input, though it can read answers aloud.
What is Gemma 4 good at on a phone? #
Strong at:
- Other languages. Gemma was trained on a very wide language mix, which makes it a good choice for translating, explaining phrases and chatting in languages other than English.
- Writing help. Rewriting, shortening, changing tone, drafting messages.
- Quick explanations of concepts, terms and instructions you paste in.
- Photos of text, like menus and signs.
Weaker at:
- Obscure facts. Small models invent answers when they don’t know; see why AI makes things up.
- Long multi-step math. Turn on thinking mode on E4B, which the catalog lists with hybrid thinking, and check the result anyway.
- Anything recent. A local model has no web access and knows nothing after its training data.
Gemma 4 or Qwen 3.5? #
Both families are in the Personal LLM catalog and you can switch between them mid-conversation, so this isn’t a decision you’re locked into. The short version: Gemma 4 E2B is the lighter, faster option on a modest phone, and Qwen 3.5 4B or 9B gives the strongest answers your RAM can support. Our head-to-head, Qwen 3.5 vs Gemma 4 on your phone, compares sizes, languages, vision and speed in detail, and how to run Qwen 3.5 is the equivalent setup guide.
A sensible approach: start with the app’s recommended first download, add Gemma 4 E2B as a fast second model, and ask both the same three questions you actually care about.
Is running Gemma 4 on your phone private? #
Yes. Gemma 4 is a file on your phone, and the answer is computed on your phone. Google provides the weights but has no connection to what you do with them locally. In Personal LLM the network is used only to download the model from Hugging Face and to show ads in the free version, and ads can’t load at all when you’re offline. There’s no account and no cloud sync of your chats.
Frequently asked questions #
Can Gemma 4 run on an iPhone? #
Yes. Gemma 4 E2B runs on iPhones with about 4 GB of RAM or more, and E4B is comfortable on iPhones with 6 GB or more. The app uses the iPhone’s GPU through Metal.
Does Gemma 4 work offline? #
Yes, once it’s downloaded. The model runs entirely on the phone, so it works in airplane mode, on a plane, or anywhere without a signal.
Is Gemma 4 free? #
Yes. Google releases the weights under the Apache 2.0 license, so they’re free to download and run, and Personal LLM is free to use with ads.
What’s the difference between Gemma 4 E2B and E4B? #
E4B is larger and gives better answers, especially with thinking mode on harder questions. E2B is smaller, faster and fits phones with less RAM. If your phone has 6 GB or more, try both and keep the one whose balance of speed and quality you prefer.
Can my phone run the bigger Gemma 4 models? #
Realistically, no. The large Gemma 4 models need far more memory than any phone provides. The mid-size one is only plausible on a 12 GB phone, loaded as your own GGUF file, and it will be slow. E2B and E4B are the phone models.