When the phone is the computer

Designing your setup around the actual constraint

For a very large share of the people reading this, the phone is not the second device. It is the only one, it has 3 or 4 GB of RAM, storage is nearly full, and data is bought in bundles. Almost every article about choosing AI tools is written by somebody for whom none of that is true.

Three constraints matter, and each has a specific answer.

Storage: do not install anything

Every major assistant works fully in a mobile browser. Open the site, then use Add to Home Screen: the browser’s install option. You get an icon and a full-screen window, using a few hundred kilobytes instead of the 200 to 400 MB a native app takes.

You also sidestep two other things: app-store pricing, which is usually higher for the same subscription, and app-store availability, which varies by country. The web version is often ahead on features anyway.

Data: know what a conversation actually costs

Text is small. This is the good news and it is worth being concrete, because people ration text chat as though it were video.

a long reply       ~= 4,000 characters  ~= 4 KB
plus page overhead ~= 20 KB per exchange, roughly
250 exchanges      ~= 5 MB

Five megabytes for a fortnight of heavy text use. Against that:

one photograph uploaded    1-4 MB   (one message = a fortnight of text)
voice mode                 ~1 MB per minute
a video call with a model  far more
a 4 GB local model         one download, then nothing

The practical rules follow directly. Text is effectively free; use it without counting. Resize photographs before uploading: most phone galleries have this, and a 1024-pixel-wide image is enough for the model to read a document while being a tenth the size. Avoid voice mode on a metered connection. And if the site keeps reloading and re-sending on a poor connection, that is the real cost, not the conversation.

An unreliable connection

Long answers stream token by token, and a connection that drops mid-answer loses the answer. Two habits help: ask for shorter replies, and copy anything valuable out immediately rather than leaving it in a tab.

If you are on a poor connection, prefer the smaller and faster model even when the strong one is available. A reply that arrives in two seconds survives a connection that a forty-second thinking-mode reply does not.

Running a model on a phone

It is possible, and the honest summary is that it works better than people expect and is more limited than the demonstrations suggest.

PocketPal and MLC Chat are free apps that run small open models on Android and iOS. A 1 to 3 billion parameter model at 4-bit quantisation is roughly 700 MB to 2 GB and runs on a mid-range phone at a usable speed.

What it is good for: drafting and rewriting, summarising text you paste in, simple translation, and working with no connection at all. What it is not good for: facts about the world, anything long, anything requiring careful reasoning. It will also warm the phone and drain the battery quickly.

The genuine use case is offline. On a train, in a field, during an outage, in a place with no coverage: a 2 GB file on your phone answers questions when nothing else will.

The cheapest useful habit on a small phone

Keep one conversation per task rather than one long conversation for everything. On a phone this is not about tidiness. A conversation that has run to two hundred messages re-sends its whole history each turn, which on a metered connection and a slow processor is felt as the app becoming sluggish and the replies arriving late.

Starting a fresh conversation, with a two-line summary pasted in if you need continuity, makes the app faster and the answers better at the same time, because the model is no longer reading three hours of earlier confusion before it reaches your question.

Other routes worth knowing

In several countries, assistants are reachable over WhatsApp or by SMS, which matters where a browser session is expensive and messaging is zero-rated on the network. These are usually feature-limited compared with the app, and worth knowing about anyway.

Shared devices deserve one line of caution. If the phone is shared with family, conversation history is visible to whoever picks it up. Use the private browsing window, or clear conversations, or keep anything sensitive off the device entirely. This is a real consideration for a large number of people and it is rarely mentioned.