Split screen restores speed in an old Galaxy’s local-AI test
A local-model experiment on an old Galaxy S23 FE ran into a problem beyond memory: response generation slowed when its server moved into the background. XDA’s Shekhar Vaidya found that keeping the server and browser visible together improved speed. The measurements describe one phone and its software settings, with the underlying scheduling mechanism still uncertain.
Artificial Intelligence··Morning
Two visible windows improve generation
In XDA writer Shekhar Vaidya’s phone test, the Chrome browser shared the display with Termux, an Android terminal environment. In that layout, Gemma’s measured speed was between 5.4 and 6.6 tokens per second. Processor observations differed too: eight cores were available to the foreground application, but the background server could use only three. Vaidya considers scheduling a possible explanation, without resolving the roles of Android and Samsung’s software. In his chosen tasks, Gemma proved more dependable; some Qwen answers contained invented dates and names of models.[1]
A small model meets Android’s background limits
XDA writer Shekhar Vaidya used an old Samsung Galaxy S23 FE to run small language models locally for summaries and email replies. His October 24 experiment began with PocketPal, an Android application for local models. The smaller Gemma 3 1B ran successfully. Attempts with Gemma 4 E2B brought memory warnings and repeated crashes. He then loaded Gemma through llama.cpp, software for running language models, inside the Termux terminal environment.[1]
The browser takes the foreground
When the browser moved to the foreground, the model server slowed sharply: the wait reached 11 minutes and 56 seconds for a response of 68 tokens. Tokens are the pieces of text a model processes and generates. Battery restrictions for Termux were already lifted, with power saving and unused-app sleeping switched off. The Qwen3.5 2B model showed the same slowdown. Its short comparison measured 0.17 tokens per second there, against 13.33 in the foreground.[1]
Related columns
For more information on this topic, you can read the related columns.
