TL;DR

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

An XDA writer tested local language models on an Exynos Galaxy S23 FE with 8 GB of RAM. Models ran, but Termux generation dropped to about 0.1 tokens per second when the app was in the background; keeping Termux visible raised Gemma 4 E2B generation to 11.39 tokens per second in a benchmark.

An XDA Developers writer found that an Exynos Galaxy S23 FE with 8 GB of RAM could run local language models, but its performance varied sharply with the app setup: Gemma 4 E2B generated at about 0.1 tokens per second through Termux when the app was in the background, compared with 11.39 tokens per second in a benchmark with Termux visible. The test suggests that running a model on an older phone may be possible, but background restrictions can make it impractical.

Shekhar Vaidya, a computing writer at XDA Developers, tested whether he could move some everyday AI tasks from cloud services to a local model on his Galaxy S23 FE. His device used Samsung’s Exynos 2200 chip and had 8 GB of RAM, of which Android showed 7.3 GB as available. Vaidya said 4.6 GB was already in use before he started testing, leaving memory as an early constraint.

He first tried PocketPal, a mobile app that offers model downloads and a chat interface. Its model list included Gemma 3 1B and Gemma 4 E2B. Gemma 3 1B loaded, but PocketPal repeatedly crashed when he tried Gemma 4 E2B after receiving a memory warning. He then used llama.cpp through Termux to load an Unsloth Q4_K_M build of Gemma 4 E2B, listed at 2.8 GB on Hugging Face.

With Termux in the background, a short Gemma test generated at around 0.11 tokens per second; a longer Qwen test was similarly slow. But when Vaidya ran benchmarks with Termux in the foreground, Gemma 4 E2B reached 11.39 tokens per second for generation. Qwen 3.5 2B reached 13.33 tokens per second on a short foreground test and about 9.98 tokens per second on a longer one. PocketPal also ran Qwen 3.5 2B, with its built-in benchmark reporting 54.54 tokens per second for prompt processing and 11.14 for generation; Vaidya’s practical tests were around 9 tokens per second.

Foreground Apps Changed Model Speed

The results show why a phone’s ability to load a model does not by itself establish that it can support a useful local-AI workflow. On this Galaxy S23 FE, the same Termux setup produced radically different benchmark speeds depending on whether the app was visible. A model that takes minutes to produce a short reply is not a practical substitute for a cloud chatbot, even if it loads successfully.

For readers, the test also separates two questions that are often conflated: whether a model fits in memory and whether the phone can run it at a workable speed. The experiment suggests that foreground execution matters on this device. It does not establish that every Samsung phone, model, or app will behave the same way, nor that the benchmark figures translate directly into everyday performance.

Answer quality was another limitation. Vaidya reported that his tests with Qwen 3.5 2B produced frequent hallucinations and invented facts. That is his assessment from the tasks he tried, not a controlled quality comparison. The findings indicate that speed alone does not determine whether a small local model is suitable for email or summarization.

Testing Two Ways to Run Models

Vaidya said he already used Gemma 4 E4B on his PC and wanted a local option for times when the computer was off. On the phone, he compared a dedicated chat application, PocketPal, with a command-line route using Termux and llama.cpp. The former was simpler to use, while Termux offered a way to run a local server and benchmark the models.

The first Termux tests appeared to show extremely poor performance. Vaidya checked the phone’s power settings: Termux was set to unrestricted battery use, unused-app sleep was disabled, and power saving was off. He then varied whether Termux was in the foreground. In his inspection of CPU availability, he found that all eight listed cores were available while Termux was active, while only cores 0–2 were available with it in the background. This observation led him to suspect that Android or Samsung’s software was limiting background CPU access.

The source does not establish that explanation as a general rule. It reports a single phone’s behavior and benchmark runs, using particular quantized model builds and settings. The measured token rates are tied to those test conditions and should not be treated as universal performance figures for the S23 FE.

“I knew it could run an LLM, so my real question was whether I could move some of the daily usage to it.”

— Shekhar Vaidya, XDA Developers

Limits of One Phone Test

The cause of the background slowdown is not independently confirmed. Vaidya observed fewer available CPU cores when Termux was not visible and proposed that Android background restrictions, possibly combined with Samsung’s software, contributed to the performance drop. The supplied account ends while discussing that hypothesis and does not report a final diagnosis or a fix.

It is also unclear how results would change across other Galaxy S23 FE variants, Android or One UI versions, model quantizations, and llama.cpp settings. The source gives benchmark results for specific runs, but does not provide enough information to treat them as controlled comparisons across all conditions. PocketPal’s built-in benchmark and Termux’s llama-bench results also use different test methods, so their numbers should not be directly compared as if they were identical measurements.

Further Tests Could Pinpoint Limits

The immediate next step would be to repeat the benchmarks under controlled conditions, varying only whether Termux is in the foreground, while recording the Android and One UI versions, CPU availability, temperature, and model settings. Testing after the phone warms up and across longer sessions would help show whether thermal behavior changes the results.

Until those details are available, the practical takeaway from Vaidya’s report is narrow: local models did run on his S23 FE, and keeping Termux visible coincided with much faster generation in his tests. Readers considering the same setup should treat the figures as device- and configuration-specific, and should check both response quality and performance for their own intended tasks.

Key Questions

Did the Galaxy S23 FE run a local language model?

Yes. XDA’s Shekhar Vaidya loaded Gemma 4 E2B and Qwen 3.5 2B on an Exynos Galaxy S23 FE with 8 GB of RAM, using PocketPal and Termux.

Why was Termux so slow in the first test?

Vaidya measured about 0.1 tokens per second with Termux in the background. He later found fewer CPU cores available in that state and suspected background restrictions, but the cause was not confirmed.

How fast was Gemma 4 E2B with Termux visible?

In Vaidya’s short foreground benchmark, Gemma 4 E2B generated at 11.39 tokens per second. That figure describes his test conditions, not a guaranteed speed for every S23 FE.

Did the local models give reliable answers?

Not consistently in Vaidya’s assessment. He reported frequent hallucinations and invented facts in his tests of Qwen 3.5 2B, so usable speed did not by itself make the model a dependable replacement for cloud AI.

Source: XDA Developers