# What AI can really run on your phone in 2026?

Date: 2026-09-02 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/local-ai-on-phone/

**Answer:** A recent phone can run models of about 2 to 4 billion parameters at a useful speed, enough for summaries, rewriting, dictation and short questions. You can use them through the system's built-in AI on newer iPhones and Android flagships, or try open models such as Gemma 4 in a free app that works offline.

Phones have become a real place to run language models. The limits are tighter than on a computer: 12 GB of shared memory is a generous amount on a phone, the battery is small, and there is no fan. So phone models are small, heavily compressed and tuned for specific jobs.

There are two ways in. The operating system may include a model that apps can call, or you can install an app that downloads an open model of your choice. They give different things, and the [private cloud and on-device guide](/en/learn/private-cloud-on-device-ai/) looks at the privacy side of the system approach.

## The models built into your phone

**On iPhone,** Apple's third generation of its Foundation Models, announced in June 2026, has two on-device members. [AFM 3 Core](https://machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models) is a 3-billion-parameter model for everyday tasks. AFM 3 Core Advanced has 20 billion parameters, but its design is unusual: the full model sits in flash storage, and only 1 to 4 billion parameters are loaded into working memory at a time, chosen for each request. Apple's [support page](https://support.apple.com/en-us/121115) lists iPhone 15 Pro, iPhone 16 and later for Apple Intelligence under iOS 27, needing up to 8 GB of storage, or up to 14 GB on iPhone 17 Pro models, iPhone Air and iPhone Duo.

Developers reach these through Apple's [Foundation Models framework](https://developer.apple.com/documentation/foundationmodels), which offers on-device models for tasks like summarizing and extracting, and points to Private Cloud Compute for harder reasoning. So not everything Apple Intelligence does stays on the phone.

**On Android,** Google's model is Gemini Nano, which runs in a system service called [AICore](https://developer.android.com/ai/gemini-nano). Google's developer documentation lists support by version:

| Version | Examples of supported phones |
|---|---|
| nano-v4 | Pixel 11 family; Galaxy Z Flip8, Z Fold8, Z Fold8 Ultra |
| nano-v3 | Pixel 9 and 10 families; Galaxy S26 family; OnePlus 15 |
| nano-v2 | Galaxy Z Fold7; OnePlus 13; Xiaomi 15 |

Gemini Nano 4 is [based on Gemma 4](https://android-developers.googleblog.com/2026/04/AI-Core-Developer-Preview.html), which comes in two sizes for phones: E2B for speed and E4B for harder tasks. Secondary reports say Nano 4 needs 12 GB of RAM. Google's [Pixel 11 specifications](https://store.google.com/product/pixel_11_specs?hl=en-US) do list 12 GB, but that requirement remains a report, not a rule Google's documentation spells out.

## What the numbers look like

Google publishes reproducible figures for Gemma 4 E2B, with the test settings, on its model card. Each run used a 2,048-token context, 1,024 tokens of input and 256 tokens of output:

| Phone | Processor used | Writing speed | Time to first word | Memory used |
|---|---|---|---|---|
| Galaxy S26 Ultra | CPU | 46.9 tokens/s | 1.8 s | 1,733 MB |
| Galaxy S26 Ultra | GPU | 52.1 tokens/s | 0.3 s | 676 MB of CPU memory |
| iPhone 17 Pro | CPU | 25.0 tokens/s | 1.9 s | 607 MB |
| iPhone 17 Pro | GPU | 56.5 tokens/s | 0.3 s | 1,450 MB |

The file is 2,583 MB. That is small because it mixes 2-, 4- and 8-bit weights, a more aggressive version of the compression explained in the [quantization guide](/en/learn/quantization-explained-gguf-mxfp4/). The model card also notes that the model loads its vision and audio parts only when needed, to save memory.

For scale, 50 tokens per second is around 35 to 40 words per second, faster than you can read. Note the 2,048-token context, however. That is about 1,500 words. The model supports up to 32,000 tokens, but more context means more memory and more heat, and a long document may behave quite differently from this test.

## A worked example: trying one offline

[Google AI Edge Gallery](https://github.com/google-ai-edge/gallery) is a free app for Android 12 or iOS 17 and later that runs open models entirely on the device. Google labels it an experimental beta.

1. Install the app from the Play Store or App Store.
2. Download Gemma 4 E2B over Wi-Fi: it is about 2.6 GB.
3. Turn on airplane mode, so you know that nothing leaves the phone.
4. Paste in a short text and try the prompt below.
5. Use the app's benchmark tool to see your own speed.

```prompt
Rewrite this message so it is polite and shorter, and keep every date and name exactly as written:

[paste your message]
```

Then test the limit. Ask it a factual question whose answer you know, such as a detail about your town. The guide on [small language models](/en/learn/small-language-models-they-enough/) explains where they fall short. Short rewriting tends to work well, while general facts are much less reliable.

French speakers can try the same test in French. Google says Gemma 4 handles more than 140 languages, and Apple's page lists French among those supported for many features, but quality in any one language is something to check yourself.

> **Try it:** Run the same paragraph through the on-device model on airplane mode and through your usual cloud chatbot. Compare the two answers on accuracy and on tone, and note how warm the phone gets after ten requests.

## Can a phone replace a laptop for private AI?

For short, personal tasks, it already can: cleaning up a message, summarizing a note, translating a sign, dictating. For anything with real length or complexity, not yet. A laptop or desktop with 32 GB can run a [27-billion-parameter coding model](/en/learn/open-coding-models-run-at-home/), roughly seven to thirteen times the parameter count of these phone models, and it can keep going without overheating.

Privacy on a phone also needs a little care. An on-device model keeps the text from being sent for that request, and Google says AICore has no direct internet access and isolates each request. But the app around the model can still save or send what you type, so read its permissions. When Apple Intelligence decides a request needs the cloud, it moves to Private Cloud Compute. A phone is a good private tool for small jobs, as long as you know which mode each app is using.

## Sources

- Introducing the Third Generation of Apple's Foundation Models: https://machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models
- ML Kit GenAI APIs: https://developers.google.com/ml-kit/genai
- Gemma 4 E2B for LiteRT-LM (model card): https://huggingface.co/litert-community/gemma-4-E2B-it-litert-lm

---

From the Guide to Artificial Intelligence, by Robert Scales, published by Holistic Hybrid Inc. Share and adapt it under CC BY-NC-SA 4.0 (https://creativecommons.org/licenses/by-nc-sa/4.0/): credit the guide, link to https://guidetoartificialintelligence.com/en/learn/local-ai-on-phone/, no commercial use without permission, and keep the same licence.