Gemma 4 Download for Android
Gemma 4's edge models run entirely on your phone - no account, no network, no data leaving the device. Here is the quickest way to install it, which size to pick, and what to expect from real hardware.
Android 12 or later · free app from Google · checked October 7, 2026
What "on your phone" actually means
Only the two edge sizes, E2B and E4B, are built for phones. The larger Gemma 4 models need a computer.
Genuinely offline
Once the model file is downloaded, it needs no network. Test it in airplane mode.
It costs battery
Sustained generation drains the battery and warms the phone faster than normal use. Idle cost is zero.
Large downloads
2.58 GB for E2B and 3.66 GB for E4B. Download on Wi-Fi and keep storage free.
A small model
Good at summaries, questions and drafting; weaker than big models on hard reasoning. See limitations.
Install with Google AI Edge Gallery
Google's free, open-source app is the easiest way to run Gemma 4 on Android. It has passed a million downloads on Google Play and was last updated there in July 2026.
Install the app
Get Google AI Edge Gallery on Google Play. Prefer a direct file? Google publishes the APK on GitHub releases.
Download a model on Wi-Fi
Open the app, choose Gemma 4 E2B or E4B, and download it inside the app.
Switch the backend to GPU
In the model settings, choose GPU. It is the biggest single speed-up - see the benchmarks.
Test in airplane mode
Turn the network off and ask a question. If it answers, everything is running on the phone.
What the app can do
AI Chat with a thinking mode, Ask Image for photos, Audio Scribe for transcription and translation, Prompt Lab, Agent Skills, Mobile Actions, and built-in model management and benchmarks.
Privacy
Prompts are processed on the device. Google Play's data-safety listing says the app collects app activity and performance data and shares none with third parties.
Bring your own model
The July 2026 update added importing LiteRT-LM models from Hugging Face community links.
On iPhone the same app is on the App Store - see mobile downloads. On a computer, start at the download hub.
E2B or E4B?
Start with E2B. Move to E4B if your phone is recent and you want better answers more than faster ones.
| Gemma 4 E2B | Gemma 4 E4B | |
|---|---|---|
| Download | 2.58 GB | 3.66 GB |
| Speed on GPU | 52 tokens per second | 22 tokens per second |
| Time to first word | 0.3 s | 0.8 s |
| Memory on GPU | 676 MB | 710 MB |
| Context | 128K | 128K |
| Inputs | Text, image, audio | Text, image, audio |
| Answer quality | Good for everyday tasks | Noticeably stronger |
| Best for | Most phones; quick replies | Recent flagships; harder questions |
Speeds measured by Google on a Samsung Galaxy S26 Ultra; your phone will differ. Quality differences are in the benchmarks; full details of both models in technical specs.
Turn on the GPU backend
It is the single biggest difference on this page. On the same phone and the same model, moving from CPU to GPU cut the wait for the first word from 1.8 seconds to 0.3, and memory use from 1.7 GB to 676 MB.
| Device and backend | Model | Reading speed | Writing speed | First word | Memory |
|---|---|---|---|---|---|
| Galaxy S26 Ultra · GPU | E2B | 3,808 t/s | 52.1 t/s | 0.3 s | 676 MB |
| Galaxy S26 Ultra · CPU | E2B | 557 t/s | 46.9 t/s | 1.8 s | 1,733 MB |
| Qualcomm NPU | E2B | 3,747 t/s | 31.7 t/s | 0.3 s | 1,869 MB |
| Galaxy S26 Ultra · GPU | E4B | 1,293 t/s | 22.1 t/s | 0.8 s | 710 MB |
| Galaxy S26 Ultra · CPU | E4B | 195 t/s | 17.7 t/s | 5.3 s | 3,283 MB |
"Reading" is prefill and "writing" is decode, in tokens per second, with 1,024 tokens in and 256 out. From the E2B and E4B LiteRT-LM model cards and Google's LiteRT-LM overview. The NPU build is a separate, slightly larger file (2,967 MB).
What your phone needs
Android 12 or later
The requirement Google lists for AI Edge Gallery.
Free storage
Room for the model - 2.58 GB or 3.66 GB - plus the app. Aim for about 6 GB free.
Memory
E2B used 676 MB on GPU and 1.7 GB on CPU in Google's tests; E4B up to 3.3 GB on CPU. More RAM gives more headroom.
A recent chip helps
GPU and NPU acceleration make the difference. Older phones fall back to CPU and run much slower.
The storage suggestion is this site's guidance, not a Google figure. For computers see the VRAM guide.
Other ways to run it on Android
AI Edge Gallery uses Google's .litertlm format, which unlocks GPU and NPU acceleration. Community apps mostly use
GGUF files from the llama.cpp world: more models to choose from, but usually CPU-bound on Android.
| App | Format | Best for |
|---|---|---|
| Google AI Edge Gallery | .litertlm | Easiest setup and the best acceleration |
| PocketPal AI | GGUF | Swapping between models; custom prompts |
| AnythingLLM | GGUF or an API | Chatting with your own documents |
| Termux + Ollama | GGUF | Command-line users; trying new models early |
The three community options are carried over from this page's earlier version and were not re-tested for this update. For GGUF files see the QAT builds; for an explanation of formats see the download hub. You can also skip installing anything and use a hosted API or the playground from your phone's browser.
Put Gemma 4 inside your Android app
Two routes: ship the model yourself with LiteRT-LM, or call the copy built into the system through AICore.
LiteRT-LM - you ship the model
Google's production runtime for on-device models, with CPU, GPU and NPU acceleration on Android and a stable Kotlin API. You control the model and version.
- E2B · 2.58 GBgemma-4-E2B-it-litert-lm
- E4B · 3.66 GBgemma-4-E4B-it-litert-lm
- Kotlin guideLiteRT-LM on Android
- Sourcegoogle-ai-edge/LiteRT-LM
AICore - the system provides it
Gemma 4 E2B and E4B are in the AICore Developer Preview, reached through the ML Kit GenAI Prompt API. Google says code written for it will work on Gemini Nano 4 devices arriving later in 2026.
- AnnouncementApril 2, 2026
- Sign upAICore Developer Preview
- AcceleratorsGoogle, MediaTek, Qualcomm
- StatusDeveloper preview
# Try the same model file on a computer first
pip install -U litert-lm
litert-lm run --from-huggingface-repo=litert-community/gemma-4-E2B-it-litert-lm --prompt="Write me a poem"
// ML Kit GenAI Prompt API - AICore Developer Preview val previewFullConfig = generationConfig { modelConfig = ModelConfig { releaseTrack = ModelReleaseTrack.PREVIEW preference = ModelPreference.FULL } } val previewModel = GenerativeModel.getClient(previewFullConfig) val previewModelStatus = previewModel.checkStatus() if (previewModelStatus == FeatureStatus.AVAILABLE) { val response = previewModel.generateContent("Summarise this note in one sentence.") } else { // Handle the case where the preview model is not available }
Command from the LiteRT-LM model card; Kotlin adapted from Google's AICore announcement, which listed tool calling, structured output, system prompts and thinking mode as coming later. Google also publishes mobile-optimised QAT weights for both sizes - see the models page.
New: EmbeddingGemma 2 on-device
Text only · 165 MB
Text and vision · 388 MB
Full model · 485 MB
Released October 6, 2026, EmbeddingGemma 2 adds on-device search across text, images, audio and video; Google's LiteRT-LM overview lists these three builds. More in this site's EmbeddingGemma 2 guide and Google's developer guide.
Is there an official Gemma 4 app on Google Play?
The official route is Google AI Edge Gallery, a free app from Google that downloads and runs Gemma 4 on your phone. There is no separate consumer "Gemma" chatbot app with account sync.
What are the minimum requirements?
Android 12 or later, and enough free storage for the model: 2.58 GB for E2B or 3.66 GB for E4B, plus the app.
Does it work offline?
Yes, completely, once the model is downloaded. Turn on airplane mode to check.
Is my data private?
Prompts are processed on the device. The app's Google Play listing says it collects app activity and performance data; read the data-safety section there before using it for anything sensitive. See also safety.
Why is it slower or weaker than Gemini?
E2B and E4B run on phone hardware; Gemini runs in data centres. The trade is that nothing is uploaded and it works with no connection. The difference is covered in the model comparison.
Does CPU or GPU matter?
Enormously. On Google's test phone, GPU cut the time to the first word from 1.8 seconds to 0.3 and memory from 1.7 GB to 676 MB.
Can I run the larger Gemma 4 models on a phone?
No. There are no mobile builds beyond E4B; the 12B needs about 6.7 GB of memory for the weights alone and the 31B about 17.5 GB. Run those on a computer - see install or the runtime selector.
Does it work on iPhone?
Yes, AI Edge Gallery is also on the App Store for iOS 17 and later. See mobile downloads.
Will it drain my battery?
While it is generating, yes - it is demanding work. When idle it uses nothing. The app also throttles between CPU and GPU depending on battery and temperature.
A capable model, offline, in your pocket
Install the app, download a model once, and it is yours to use anywhere.