Qwen3 0.6B on a 4 GB Galaxy A05s: 14.8 tok/s.
Our small open model, taught only one app's tasks, writes 14.8 tok/s on a 4 GB Galaxy A05s with llama.cpp, and 11.0 tok/s with 768 tokens already in context. FTL
Measured on a real phone in Firebase Test Lab (FTL) with the setting that shipped: f16 KV, flash-attn 0, repack 0. Network not controlled (not airplane mode). Published 2 Nov 2026.
What are the four shipped-setting rows?
Mean and standard deviation (sd) of three llama-bench repetitions, in tokens per second. Depth d0 starts from an empty context; d768 starts with 768 tokens already in context.
sd 0.16. Shown on our site as 14.8 tok/s.
sd 0.01. Shown on our site as 11.0 tok/s.
sd 1.35. Shown on our site as 25 tok/s.
sd 0.07.
Network not controlled (not airplane mode). Peak PSS 744 MB for that run. The bench harness does not set or record the network state.
Bench run conditions FTL
Galaxy A05s SM-A057G, Snapdragon 680 'bengal', 4 GB, Android 15; llama.cpp 620d222 (armv8a); qwen3-0.6b-v18-Q6_K.gguf 489 MB; 4 threads; f16 KV, flash-attn 0, repack 0; llama-bench -r 3 (mean of 3); thermal_status 0, battery 31 °C at start.
The same source 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.
Data source: the four rows, with their standard deviations and the FTL layer tag, are in our public dataset neykosa/phone-llamacpp-bench on Hugging Face (CC BY 4.0). Source run file: ftl_raw_a05s_v18q6_fa_kv_20261026.json.
How did the whole app perform as offline AI?
A separate FTL run tested the whole Nesayo offline edition, our first build, on the same phone model with airplane mode on. Answer times, model memory and first launch come from this run, not from the speed bench above.
- Test phone FTL
- Galaxy A05s, 4 GB, Snapdragon 680. Android 15, airplane mode, run PASSED
- Stock answer, offline FTL
- 4.25 s. The first turn of that run took 13.4 s (right after a print preview); still open
- Bill from a sentence FTL
- 5.76 s
- Model memory FTL
- 673 MB. The model in memory during the whole-app run (the speed bench reports its own peak PSS above)
- First launch FTL
- 18.8 s. Loads the model once (cold_start_ms 18786)
Source run file: ftl_final_whole_app_a05s_26efc64a1_20261031.json. This file is not public.
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. These rows describe the model in the Nesayo offline edition. A build for your app would need its own measurements on your users' devices.
The measured weights are private. To compare on your own phone, 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.
Questions
Can a 4 GB Android phone run an LLM offline?
On our test phone, yes. In the whole-app run with airplane mode on, a Galaxy A05s with 4 GB answered a stock question in 4.25 s, with 673 MB of model memory. FTL
Which llama.cpp settings produced 14.8 tok/s?
llama.cpp 620d222 (armv8a), 4 threads, f16 KV, flash-attn 0, repack 0, a Q6_K model of 489 MB, mean of 3 llama-bench repetitions. Network not controlled (not airplane mode). FTL
Want a small open model taught only your app's tasks?
Our design-partner offer is for software companies whose app handles data it cannot send to a cloud AI.
Tell us what your software does, one task you want the model to handle and your target devices.