TL;DR
Get the latest gadgets delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
An XDA writer tested local language models on an Exynos Galaxy S23 FE and found that Termux generation speed dropped to about 0.1 tokens per second when the app was in the background. With Termux in the foreground, the same phone reached 11.39 tokens per second on Gemma 4 E2B; the writer traced the change to fewer CPU cores being available, though the precise cause remains a hypothesis.
An XDA writer found that a Samsung Galaxy S23 FE could run local language models at usable speeds in Termux only while the app remained in the foreground. In the reported tests, Gemma 4 E2B generation rose from about 0.1 tokens per second with Termux in the background to 11.39 tokens per second in front, a difference that affects whether an older phone can serve as a practical local AI device.
The test used an Exynos 2200 Galaxy S23 FE with 8 GB of RAM, of which Android showed 7.3 GB as available. The writer first tried PocketPal, a local-model app that recommended several models for the hardware. Gemma 3 1B loaded, but Gemma 4 E2B repeatedly triggered a memory warning and caused PocketPal to crash while loading.
The writer then installed llama.cpp through Termux and loaded an Unsloth Q4_K_M build of Gemma 4 E2B, listed at 2.8 GB and about 4.65 billion parameters including embeddings. The model loaded and served through localhost, but initial browser prompts were extremely slow: an eight-token response took 1 minute 11 seconds, while a 68-token response took 11 minutes 56 seconds. The writer checked battery and power settings; Termux was set to unrestricted battery use, power saving was off, and unused-app sleep was disabled.
Further testing pointed to whether Termux was visible. A short Qwen 3.5 2B benchmark recorded 0.17 tokens per second for generation when the app was not in front and Chrome was active, compared with 13.33 tokens per second with Termux in front. A longer foreground Qwen run reached 9.98 tokens per second, and Gemma 4 E2B reached 11.39. In a separate PocketPal test, Qwen’s built-in benchmark showed 11.14 tokens per second for generation, while real prompts ran at about 9 tokens per second. The writer also reported frequent hallucinations in the model’s answers for everyday tasks.
Mobile AI / Hands-on report
Can an Old Galaxy Become a Local AI Host?
An XDA writer tested local language models on an Exynos Galaxy S23 FE. The striking result: Gemma 4 E2B generated about 100 times faster when Termux stayed in the foreground.
One phone, two very different speeds
The report’s central finding was app visibility. A short Qwen benchmark also showed a sharp foreground/background split. The chart compares the reported generation speeds on one shared scale.
Measurements came from the writer’s setup and prompts. They are not a performance guarantee across devices, software versions, thermal conditions, or model builds.
From app crash to local server
The writer moved from a simple model app to a terminal based setup after PocketPal could not load the larger Gemma model.
Why did background speed fall?
The writer observed that all eight CPU cores were available while Termux was active. In the background, the reported available list fell to cores 0–2. This points to a possible explanation, but the exact Android or Samsung mechanism was not confirmed.
Fewer cores listed
Linux process information showed a smaller available core set when Termux was not in front.
Precise cause
The report did not confirm which system policy caused the change or whether it applies across Android builds.
Background hosting
No workaround was demonstrated that preserved foreground-level speed while Termux was hidden.
Fast output does not guarantee reliable answers
The writer reported frequent hallucinations and invented facts during everyday prompts. Treat the benchmark as a speed observation, not an accuracy measure.
“I can’t always keep my PC running, so I thought of running a local LLM on my old Galaxy S23 FE.”
Shekhar Vaidya · XDA writer“Gemma 4’s 0.1 tok/s generation jumped to 11.39 tok/s.”
Shekhar Vaidya · XDA writer“Most of the generations had a lot of hallucinations and invented facts.”
Shekhar Vaidya · XDA writerWhat the test supports
Could the S23 FE run a local model?
Yes. Models loaded through Termux, and PocketPal ran Qwen 3.5 2B. Performance varied by model and app state.
How fast was Gemma 4 E2B in front?
The reported foreground generation speed was 11.39 tokens per second, compared with about 0.1 in an earlier background run.
Was the slowdown’s cause confirmed?
No. Fewer available CPU cores were observed in the background, but the precise system cause remains a hypothesis.
Was a background performance fix found?
No fix was reported. Controlled tests across prompts, app states, devices, and Android versions could clarify the limits.
Foreground Use Changes Phone Performance
The results suggest that an older phone may be capable of running a small local model, but raw memory capacity and model size are not the whole story. In this test, keeping Termux in the foreground was associated with a large change in generation speed, making the difference between waiting several minutes for a short response and getting output at a more usable pace.
That matters to people considering local AI for privacy, offline access, or reduced reliance on cloud services. The test also shows why a single benchmark can mislead: the same device and model produced sharply different results under different app states. These are one writer’s measurements, not a broad performance guarantee for every Galaxy S23 FE, model build, or Android setup. The reported hallucinations also mean speed alone does not establish that a model is dependable for practical tasks.
Samsung Galaxy S23 FE USB-C cooling fan
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
From PocketPal to Termux
The writer, Shekhar Vaidya, said he wanted to shift some routine AI use away from cloud services and had previously run Gemma 4 E4B on a PC. Because the computer was not always available, he tried using an older Galaxy S23 FE instead. The phone’s memory was already under pressure: the report says about 4.6 GB of its 8 GB was in use before the larger model test.
PocketPal offered a simple interface and a selection of models suited to the device, but it crashed when loading Gemma 4 E2B. Termux provided another route through llama.cpp, and the model started successfully. The initial slow results led the writer to suspect Android power management, but the phone’s power settings did not explain the performance gap. Subsequent tests compared foreground and background execution, and the writer inspected CPU availability through Linux process information.
““I can’t always keep my PC running, so I thought of running a local LLM on my old Galaxy S23 FE.””
— Shekhar Vaidya, XDA writer
USB-C phone cooler for Android gaming
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Why Background Speed Fell
The test reports that all eight CPU cores were available while Termux was active, whereas the available list fell to cores 0–2 when the app ran in the background. The writer suggested Android restricts apps without a visible window and noted Samsung’s software as a possible part of the explanation. That explanation remains a hypothesis; the report does not establish the exact system mechanism or show that the same restriction applies across Android versions and devices.
The results are also specific to the stated phone, model quantization, software setup, and benchmark prompts. The report does not provide a controlled comparison across multiple S23 FE units or establish whether other background workloads, thermal conditions, or model settings affected every run. It also does not identify a fix that would let Termux retain foreground-level speed while hidden.
portable USB-C hub with Ethernet for Android phone
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Further Tests Could Pinpoint Limits
The report ends with the foreground-versus-background findings and does not announce a confirmed software fix or follow-up test. A useful next step would be to repeat the benchmarks under controlled conditions, including the same prompt, model, temperature, and app state, while checking which CPU cores Android makes available. Tests on other Samsung phones and Android builds could show whether the behavior is specific to this setup.
For readers trying the same approach, the account supports a limited conclusion: local models can load on this S23 FE, and the tested Termux setup ran much faster when visible. It does not establish reliable background hosting or guarantee accurate answers. The practical question still to be resolved is whether the core restriction can be explained and avoided without keeping the terminal app open.
Samsung Galaxy S23 FE phone stand with cooling fan
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Questions
Could the Galaxy S23 FE run a local language model?
Yes. In the reported test, the Exynos Galaxy S23 FE loaded models through Termux, and PocketPal ran Qwen 3.5 2B. Performance varied considerably by app state and model.
Why did Termux perform so slowly at first?
The writer found that generation slowed to about 0.1–0.17 tokens per second when Termux was not in the foreground. They also found fewer CPU cores available in that state, but the exact reason for the restriction was not confirmed.
How fast was Gemma 4 E2B with Termux in front?
The reported short benchmark measured 11.39 tokens per second for generation with Termux active, compared with about 0.1 tokens per second in the earlier background run.
Did the local model give reliable answers?
Not consistently, according to the writer, who reported frequent hallucinations and invented facts during everyday-use testing. The speed measurements do not establish answer accuracy.
Was a fix for background performance confirmed?
No fix was reported. The writer’s CPU-core observation points to a possible reason for the slowdown, but further testing would be needed to confirm the cause and identify a way to maintain performance in the background.
Source: XDA Developers
Evergreen bestsellers Picks
bestsellers
As an affiliate, we earn on qualifying purchases.
