Module 4: The parts that matter for AI · Lesson 4.6 · about 15 min
AI on phones and tablets
What small models can do on a phone or tablet with no internet, which apps run them, and the limits of battery, heat and memory.
Module 4 · about 15 minutes · read first: 4.1. You’ll be able to put a small model on your phone and know what to expect from it.
1. Why this matters
Your phone is the device you always carry, including on flights, on trains and in places with poor signal. A small offline model there can draft, rewrite, translate and answer short questions with nothing leaving the phone. Think of it as a pocket dictionary: not the whole library, but always with you.
2. What fits
Phones in 2026 have about 8 to 16 GB of memory, shared with the system and your apps. What is left for a model is about 2 to 6 GB. That means models of roughly 1 to 4 billion parameters at 4-bit, such as Gemma 4 E2B and E4B, which Google designed for phones and laptops. (The “E” stands for effective parameters: the model runs in about the memory of a 2B or 4B model, even though it holds more parameters in total, because some of them can stay out of fast memory.) Phone makers also build in their own small on-device models for features such as summaries and smart replies.
Here is the sum for a typical phone:
- GB for weights
- What the phone can spare, minus a little for the conversation.
- × 8 ÷ bits
- Turns GB into billions of parameters: 8 bits per byte, divided by the bits each weight uses.
- Effective parameters (E2B, E4B)
- Phone models that run in about the memory of a 2B or 4B model, though they hold more parameters in total.
- Physical meaning
- A phone has a small desk and no fan. The model must fit on that small desk, and the work must be short enough that the phone does not overheat and slow down.
3. How to try it
- Android: Google AI Edge Gallery (open source) downloads Gemma models and runs them offline.
- iPhone and iPad: apps from the App Store that run open models on the device. Check that the app says the model runs on the device and works in flight mode.
- Download the model on Wi-Fi first (phone versions are usually 1.5 to 4 GB), then switch on flight mode to test (Lesson 9.6).
“If an app has AI in it, it works offline.” Many phone AI apps send your words to a server in the cloud. Only apps that download a model to the phone can work with no signal. The flight-mode test settles it: if it still answers with every connection off, it is truly on the device.
4. Limits
- Speed: often 10 to 30 tokens per second for the smallest models on a recent phone; older phones are slower.
- Heat and battery: phones have no fan. Long sessions warm the phone, then it slows itself down. A long session can use a noticeable share of the battery.
- Context: keep prompts short; long documents eat the little memory left.
- Quality: small models make more mistakes (Lesson 2.7). Check facts.
Maths and principle
Principle: a phone is a tiny-memory, no-fan computer: small models, short jobs.
Equation: largest model ≈ usable GB × 8 ÷ bits per weight. With 3 GB at 4.8 bits: 3 × 8 ÷ 4.8 = 5 billion parameters at most, before context.
Analogy: a pocket dictionary. Not the whole library, but always with you.
Try it: your tablet can spare 6 GB. Largest model at 4.8 bits? (6 × 8 ÷ 4.8 = 10B, but leave room for context: realistically 4B to 8B.)
5. Worked example
Made-up numbers. Leena travels by train through areas with no signal. She installs an offline app with Gemma 4 E2B on Wi-Fi at home. On the train she rewrites emails and translates a few menus; each answer arrives in a few seconds. After 20 minutes of steady use the phone is warm and slower, so she saves long jobs for her laptop.
6. Practice
- How much memory does your phone have? (Settings > About.)
- Install an offline model app, switch on flight mode and ask it to rewrite a paragraph.
- Why should you not paste a 40-page document into a phone model?
Answers
- Note it; usable memory for a model is roughly a third of it.
- If it answers in flight mode, it is truly offline.
- The context would not fit in the little memory left, and the phone would heat and slow down.
New words: on-device, effective parameters (the E in E2B), flight-mode test.
- A phone can usually spare 2 to 6 GB for a model: enough for about 1B to 4B parameters at 4-bit.
- Largest model (billions) ≈ GB for weights × 8 ÷ bits per weight; pick something smaller to leave room.
- Download on Wi-Fi, then use the flight-mode test to prove the app really runs on the device.
- No fan means heat and slowdown on long jobs: keep phone jobs short and check facts.
Check yourself
Sources
- Gemma 4 small models (E2B, E4B) built for phones and laptops: Google, Gemma model card (official page), https://ai.google.dev/gemma
- Google AI Edge Gallery, an app that runs open models on Android offline (official page), https://github.com/google-ai-edge/gallery
- Dettmers, T., Lewis, M., Belkada, Y. and Zettlemoyer, L. (2022), LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale (why fewer bits per weight shrink memory), arXiv:2208.07339 https://arxiv.org/abs/2208.07339
Your ticks stay in this browser. Save them as a file on the course home.
Updated 2026-11-22. Prices and model names change fast: check the official source before you buy.