Team Neykosa measured a 0.6B model writing 14.8 tokens a second on a 4 GB Galaxy A05s. But a user never waits for tokens a second. They wait for an answer, so we report the two separately.
The speed bench
The model is Qwen3 0.6B, a small open model taught only one app's tasks, at 6-bit Q6_K. It ran with llama.cpp on a real Galaxy A05s (SM-A057G, Snapdragon 680, 4 GB, Android 15) in Firebase Test Lab (FTL). Settings: llama.cpp 620d222 (armv8a), 4 threads, f16 KV, flash-attn 0, repack 0. Each row is the mean of three llama-bench repetitions, in tokens per second (layer FTL):
| Test | Context depth | tok/s | sd |
|---|---|---|---|
| tg64, writing | 0 | 14.82 | 0.16 |
| tg64, writing | 768 | 11.02 | 0.01 |
| pp128, prompt reading | 0 | 25.45 | 1.35 |
| pp128, prompt reading | 768 | 20.33 | 0.07 |
Network not controlled (not airplane mode): the bench harness does not set or record the network state. Peak PSS was 744 MB for that run. The same run also measured flash-attn on, q8_0 KV and weight repack on. Those settings did not ship, so their rows are left out here and in the public dataset.
Response times come from a different run
A separate FTL run tested a test build (31 Oct) of the whole Nesayo offline edition, our first build (pharmacy software, nesayo.com), on the same phone model with airplane mode on:
- Stock answer, offline: 4.25 s. The first turn of that run took 13.4 s, right after a print preview; that gap is still open.
- Bill from a sentence: 5.76 s.
- Model memory: 673 MB (the speed bench reports its own peak PSS above).
- First launch: 18.8 s.
No later merged build has passed a whole-app run yet. The live Nesayo product uses cloud AI; this offline edition is tested on real phones in a device lab, and a shop pilot is next.
What these numbers do not show
- Answer quality. Speed rows say nothing about whether an answer is right.
- Battery or energy use. Energy use was not measured.
- Other phones. These are results from one phone model, the Galaxy A05s.
- Your app. A model for a different app would need its own measurements on its users' devices.
The data
The four speed rows, with their standard deviations and the FTL layer tag, are in the public dataset neykosa/phone-llamacpp-bench on Hugging Face (CC BY 4.0). The measured weights are private. The dataset card gives a recipe with the public Qwen/Qwen3-0.6B checkpoint at the same settings; matching the settings does not reproduce our model.
If you have a llama-bench result from another phone, the dataset's Discussions tab is where to post it, with whatever settings you recorded.
The full page, with the method and a short Q&A: neykosa.com/benchmarks.
This article was written by an AI agent working for Team Neykosa and checked against our published claims list before posting.

