Run AI Offline on Your Phone in India: No Data Pack, No Subscription (2026 Guide)
How we wrote this guide: we did not test these models on our own phones. Everything here comes from Google's published benchmarks and documentation, the official model cards and published hands-on reviews. Every figure links to its source at the end, and where sources disagree or say nothing, we tell you.
Short answer: yes, you can now run a genuinely useful AI model on your phone with no internet at all. Install Google's free AI Edge Gallery app, download the Gemma 4 E2B model once on Wi-Fi (about 2.6 GB), then switch on airplane mode. It chats, reads photos and transcribes audio entirely on the phone. You need Android 12 or iOS 17 or later, and in practice a phone with about 6 GB of RAM for E2B and about 8 GB for the stronger E4B model.
For most Indians, AI still means an app that needs mobile data, a login and often a subscription. Offline AI changes three things at once. It works on a train through a tunnel or in a village with one bar of signal. It uses no data after the first download. And nothing you type or photograph leaves your phone.
What "offline AI" means
ChatGPT, Gemini and Claude run on huge servers. Your phone sends your message over the internet and the answer comes back. Offline AI is different: a smaller model is stored on your phone and runs on its own processor.
The models that made this practical are Google's Gemma 4 family, released in 2026 under the open Apache 2.0 licence. Two sizes are built for phones:
| Model | Effective size | Context | Understands | Phone download (LiteRT) |
|---|---|---|---|---|
| Gemma 4 E2B | 2.3B parameters (5.1B with embeddings) | 128K tokens | Text, images, audio | About 2.6 GB |
| Gemma 4 E4B | 4.5B parameters (8B with embeddings) | 128K tokens | Text, images, audio | About 3.7 GB |
The bigger Gemma 4 models (26B and 31B) need a laptop or desktop with a lot of memory. They are not phone models.
How to set it up, step by step
- Check your phone. Google's AI Edge Gallery needs Android 12 or later, or iOS 17 or later. On Android, see Settings, then About phone, for your Android version and RAM.
- Install AI Edge Gallery. It is on the Google Play Store and the Apple App Store. If Play Store is not available on your phone, Google also publishes an APK on the app's GitHub releases page.
- Connect to Wi-Fi. The model download is 2.6 GB to 3.7 GB, which is more than a typical daily mobile data pack.
- Download a model. Open the model management section and choose Gemma 4 E2B or E4B. This is the only step that needs internet.
- Switch on airplane mode and use it. A hands-on review by Analytics Vidhya confirmed the app's features kept working offline after the download.
The app is labelled an experimental beta, and Google says it runs all inference on the device.
What you can do with it
AI Edge Gallery has several modes, according to its official GitHub page:
- AI Chat: a normal chatbot, with an optional Thinking Mode for harder questions.
- Ask Image: point the camera at something or pick a photo and ask about it.
- Audio Scribe: transcribe speech and translate it.
- Prompt Lab: test a single prompt with settings such as temperature.
- Agent Skills: small tools such as Wikipedia lookup and maps.
- Mobile Actions: control some phone functions offline.
- Model Management and Benchmark: download models and measure how fast they run on your own phone.
That last feature is useful. Since we did not test on our own devices, the benchmark tool is the best way to see exactly how your phone performs.
Some realistic uses in India:
- Study on the move: ask questions about a topic or have notes explained during a long train journey with no network.
- Private documents: ask about a photo of a form or letter without uploading it anywhere.
- Weak-signal areas: get help with writing, summaries or simple maths where mobile data is unreliable.
- Saving data: after the one-time download, every question costs zero data.
Which phones can run it
Google does not publish an official minimum RAM for AI Edge Gallery. It lists only the Android and iOS versions. What we have instead are memory measurements from Google's model cards and a community guide.
- Google's published benchmarks show Gemma 4 E2B using about 0.7 GB to 1.7 GB of memory while running on a Samsung Galaxy S26 Ultra, and E4B about 0.7 GB to 3.3 GB, depending on whether the phone's GPU or CPU does the work.
- Analytics Vidhya reports E2B needs less than 1.5 GB of RAM and E4B about 2.5 GB.
- Your phone also needs memory for Android itself and your other apps, so total RAM matters more than these figures suggest.
A community guide on Gemma4All gives this rough rule, which is a sensible starting point but not an official requirement:
| Your phone's RAM | Gemma 4 E2B | Gemma 4 E4B |
|---|---|---|
| 4 GB | Not recommended | No |
| 6 GB | Maybe, can be tight | No |
| 8 GB | Yes | Yes |
| 12 GB or more | Yes | Yes |
Many mid-range phones sold in India now come in 8 GB variants, but RAM differs between variants of the same model, so check your own phone's settings rather than the model name.
How fast is it?
Google publishes speed benchmarks for both models on its LiteRT-LM model cards. "Decode" is how fast the answer is written, in tokens per second. "Time to first token" is the wait before the answer starts.
| Model | Device | Runs on | Answer speed (tokens/sec) | Wait before answer |
|---|---|---|---|---|
| E2B | Samsung Galaxy S26 Ultra | GPU | 52.1 | 0.3 s |
| E2B | Samsung Galaxy S26 Ultra | CPU | 46.9 | 1.8 s |
| E2B | iPhone 17 Pro | GPU | 56.5 | 0.3 s |
| E4B | Samsung Galaxy S26 Ultra | GPU | 22.1 | 0.8 s |
| E4B | Samsung Galaxy S26 Ultra | CPU | 17.7 | 5.3 s |
| E4B | iPhone 17 Pro | GPU | 25.1 | 0.9 s |
Google measured these with a 1,024-token input, a 256-token answer and a 2,048-token context. With a newer feature called speculative decoding, Google's E4B card shows the S26 Ultra GPU reaching 46 to 49 tokens per second on summaries, code and tone rewriting.
These are flagship phones. Google has not published numbers for budget or mid-range Android phones, so expect your phone to be slower, possibly much slower. Community reports collected by Gemma4All put recent Snapdragon flagships at roughly 8 to 20 tokens per second, but those are not controlled tests.
For a sense of scale, a common rule of thumb is that one English token is about three quarters of a word, so even 10 tokens per second writes faster than most people read. Indian-language text needs more tokens for the same words, so it will feel slower. In our own measurement, the Gemma 3 tokenizer needed 1.37x as many tokens for Hindi as for English, and 1.89x for Tamil. See does AI cost more in Hindi for the full data.
Does it work in Hindi and other Indian languages?
Here we have to be careful, because the published evidence is thin.
- Google's Gemma 4 model card says the models were pre-trained on more than 140 languages, with out-of-the-box support for 35 or more. The card does not list which languages are included, so it does not confirm Hindi or any other Indian language by name.
- On MMMLU, a multilingual knowledge test, the model card reports 76.6% for E4B and 67.4% for E2B. For comparison, last year's much larger Gemma 3 27B scored 70.7%, so the phone-sized E4B beats it.
- We could not find a published Hindi-specific benchmark for E2B or E4B. Artificial Analysis publishes a Hindi leaderboard, but it does not include these small models.
So the honest answer: Gemma 4 does well on multilingual tests overall, but Google does not name the languages on its model card and nobody has published a careful Hindi score for the phone models. Try it on your own questions in your language before relying on it, and expect it to be weaker than in English, especially for smaller languages.
The limits you should know
Offline AI on a phone is impressive, but it is not ChatGPT in your pocket.
- It is less capable. These are small models. In Analytics Vidhya's test, E2B stopped mid-task on a coding request that E4B completed.
- It does not know recent events. With no internet, it cannot look anything up. It only knows what it learned in training.
- It can be slow on older phones. The same review found offline generation slow and agent features unreliable at times.
- It drains the battery. Running a model is heavy work for a phone. Phones with a dedicated AI chip handle it better than CPU-only phones.
- Keep chats short. The models support long context on paper, but phone memory runs out sooner. Start a new chat for a new topic.
- It still makes mistakes. Like all AI models, it can state wrong facts confidently. Check anything important.
Other ways to get offline AI
- Built into some phones: Google's Gemini Nano already runs on-device on supported Android phones. In April 2026, Google said the next version, Gemini Nano 4, is based on Gemma 4 and would come to new flagship Android phones later in the year, claiming up to 4x faster responses and up to 60% less battery use than the previous version.
- Other apps: community apps such as Off Grid can load Gemma 4 and other open models in GGUF format. Check the permissions and source of any third-party app before giving it your data.
- On a laptop: if you have a laptop with 16 GB of memory or more, you can run larger models such as Gemma 4 26B or Qwen coding models. Our local versus cloud AI guide covers that.
When you still need online AI
Use offline AI for privacy, travel and saving data. Switch to an online model when you need the latest information, long documents, serious coding or the highest accuracy. If you build apps, a sensible pattern is to handle simple requests on the phone and send hard ones to a large model online.
For the online part, Velona gives developers GPT, Claude, Gemini, DeepSeek, Qwen and many more models through one OpenAI-compatible API, paid in rupees with UPI and billed only for what you use.
Frequently asked questions
Can I use AI without internet on my phone?
Yes. Install Google AI Edge Gallery, download a Gemma 4 model once on Wi-Fi, and it works in airplane mode. Chat, image questions and audio transcription all run on the phone.
Can I run ChatGPT offline?
No. ChatGPT runs only on OpenAI's servers and needs internet. Offline apps use smaller open models such as Gemma 4 instead, which are less capable than ChatGPT but work with no connection.
How much storage does offline AI need?
Google's LiteRT model files are about 2.6 GB for Gemma 4 E2B and about 3.7 GB for E4B. Keep extra free space for the app and for your phone to work smoothly.
Will it work on a 4 GB or 6 GB RAM phone?
Google does not publish a minimum RAM. Community guidance suggests about 6 GB as the practical minimum for E2B and about 8 GB for E4B. Phones with 4 GB are unlikely to run it well.
Does offline AI work in Hindi?
Gemma 4 was pre-trained on over 140 languages and scores well on multilingual tests, but Google does not list the languages by name and no Hindi-specific benchmark has been published for the phone-sized models. Test it with your own questions first.
Is offline AI private?
Google says AI Edge Gallery runs all processing on the device, so your messages and photos are not sent to a server while you use it. You still need internet once to download the app and model.
Is it free?
Yes. The AI Edge Gallery app is open source under Apache 2.0, and the Gemma 4 models are free under the same licence. The only cost is the one-time download.
Sources
- Google AI Edge Gallery (Google, GitHub)
- Gemma 4 model card (Google AI for Developers)
- Gemma 4 E2B LiteRT-LM benchmarks (Google, Hugging Face)
- Gemma 4 E4B LiteRT-LM benchmarks (Google, Hugging Face)
- Gemma 4: the new standard for local agentic intelligence on Android (Android Developers Blog, April 2026)
- How to run Gemma 4 on your phone without internet (Analytics Vidhya, April 2026)
- Run Google's new Gemma 4 AI models locally on Android and iOS (Business Today, April 2026)
- Gemma 4 on phone: run it on iPhone or Android (Gemma4All, community guide)
- Hindi language AI benchmark (Artificial Analysis)