An old-phone assistant needs room for two windows

Before the first reply

Using an old phone to draft a short email reply is an appealing idea: the computer can stay off, and a small task need not require another service. XDA writer Shekhar Vaidya follows that idea from loading screen to usable output in his Galaxy S23 FE experiment. The phone has 8 GB of memory; about 4.6 GB of the 7.3 GB visible to Android is already occupied. PocketPal, the easiest route, runs Gemma 3 1B. His preferred Gemma 4 E2B crashes after a memory warning. The first everyday obstacle is familiar: the app leaves before the task begins.^[1]

Vaidya gets past that obstacle by running the model through Termux. It loads, but moving to the browser to read the response introduces another problem. With the server in the background, a 68-token answer takes 11 minutes and 56 seconds to generate. For a short email draft, that wait could consume the convenience of handing the job to the phone. Termux was already unrestricted in the battery settings, with power saving off. After successfully loading the model, there is still the question of how to use the phone while waiting for its answer.^[1]

Making room for two windows

Screen arrangement becomes part of the job. In Vaidya’s short Qwen3.5 2B measurement, generation runs at 0.17 tokens per second with Termux in the background and 13.33 in the foreground. His CPU observations likewise show the first three cores available in the background and all eight available in the foreground. He suspects operating-system scheduling without separating Android’s contribution from Samsung’s software. That leaves no basis for presenting the same setting as a fix for every phone. What is concrete is that keeping the application visible changes the wait in this setup.^[1]

The working recovery is split screen: Termux stays on one side and Chrome on the other. Gemma generates 5.4–6.6 tokens per second in that arrangement, with a longer 730-token example finishing in about two minutes and four seconds. There is usable speed for short responses, at the cost of making room for two apps. To me, that small bargain is the most useful result. The old device can return to work, provided the server’s window stays beside the answer. What happens when another app takes over becomes an immediate usability concern.^[1]

The work left in the draft

Once speed recovers, the text still needs attention. Vaidya found Gemma more dependable in his selected summary, email and calculation examples; Qwen invented some model names and dates. That is not a general accuracy ranking of the two models. Running locally does not automatically correct false information either. For an email draft, the remaining user job is to read the names, dates and commitments and change them where necessary. Considering delay and correction together on the same small task makes the useful boundary easier to see.^[1]

For an enthusiast, the tinkering itself can have value: reusing an idle phone and finishing a small task on a device they already own. People differ in the time they can spend setting up a command line or keeping two windows visible. I therefore cannot read this experiment as evidence that every old phone is a ready assistant. The convenience I would wait for is more ordinary: opening the model and reading its reply without having to babysit a second window. On this Galaxy, the working route still keeps that window in view.^[1]