neykosa/phone-llamacpp-bench

Qwen3 0.6B on a 4 GB Galaxy A05s: 14.8 tok/s at depth 0, 11.0 tok/s at depth 768

Layer FTL: measured on a real phone in Firebase Test Lab. These are llama-bench writing speeds for Neykosa's private app-tuned Qwen3 0.6B model, quantised to Q6_K, using the shipped setting.

Context depth matters when comparing phone benchmarks. This dataset publishes the shipped-setting rows, a recipe for contributing measurements from your own phone, a script that prints your result ready to paste, and our own row written out as a sample submission. Contributions use a public comparison model and remain separate from our private-model results.

Measured rows — FTL

Source run: ftl_raw_a05s_v18q6_fa_kv_20261026.json.

The shipped setting: f16 KV, flash-attn 0, repack 0. Values are the mean and standard deviation of three llama-bench repetitions.

Test	Context depth	Mean tok/s	Standard deviation, tok/s
pp128 — prompt reading	d0	25.45	1.35
tg64 — writing	d0	14.82	0.16
pp128 — prompt reading	d768	20.33	0.07
tg64 — writing	d768	11.02	0.01

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.

Where this benchmark falls short

• One phone model. These rows say nothing about another phone.
• You cannot rerun our model. The measured weights are private. The recipe below uses the public checkpoint, so your numbers describe a different model.
• Network not controlled. Do not read these rows as an airplane-mode result.
• Context costs speed. Writing falls from 14.82 tok/s at d0 to 11.02 tok/s at d768, so a d0 number alone flatters an app that carries chat history.
• The noisiest row. pp128 at d0 has the widest spread, sd 1.35.
• Not in this dataset. Answer quality and whole-app response times. Energy use was not measured.

The whole-app run, with its own conditions and limits, is on our benchmarks page. A plain-language reading is in Can a small language model run on a budget Android phone with 4 GB RAM?

Data file

a05s_shipped_setting.csv contains the four shipped-setting rows above, their means and standard deviations, the FTL layer tag and the source run filename.

The same source file also measured flash-attn on, q8_0 KV and weight repack on. Those settings are excluded from this dataset.

This is a speed-measurement dataset: no weights, prompts, answers or user data. The measured model was taught one app's tasks. Shipped weights and customer data remain private; we publish measured results, method write-ups and synthetic-data examples.

Before you start

You need:

• A computer with adb (Android platform tools) and a USB cable.
• An Android phone with an arm64 processor and USB debugging switched on in Developer options.
• llama-bench built for Android arm64 from llama.cpp commit 620d222. llama.cpp's own guide covers the build: Cross-compile CLI using Android NDK. We have not walked an outside build through those steps, so a note on where yours failed helps the next person.
• Free phone storage for the model file, and Python on your computer for bench_row.py.

This recipe times the model from the phone's shell over adb. It does not show how to embed llama.cpp inside an app.

Run the public comparison model

The measured tuned weights are private, so use a public file for your contribution: Qwen3-0.6B-Q6_K.gguf, a community quantisation of Qwen/Qwen3-0.6B. Matching the benchmark settings does not reproduce our private model. Nothing needs converting.

Build llama-bench for Android arm64 from llama.cpp commit 620d222, then, from your computer:

adb push llama-bench Qwen3-0.6B-Q6_K.gguf /data/local/tmp/
adb shell 'cd /data/local/tmp && chmod +x llama-bench && PP=pp128 TG=tg64 && ./llama-bench -m Qwen3-0.6B-Q6_K.gguf -t 4 -r 3 -p ${PP#pp} -n ${TG#tg} -ctk f16 -ctv f16 -fa 0 --repack 0 -d 0,768 -o json' > run.json

PP and TG hold the two test names; ${PP#pp} and ${TG#tg} give the -p and -n lengths. Keep --repack 0 to match the shipped setting. The requested tests are pp128 and tg64 at depths 0 and 768.

We ran this file and these flags on an x86 build at 620d222: it loaded and wrote all 4 rows. We did not run the adb steps or time a phone with this public file for this card, so tell us where the recipe breaks.

Submit your phone's result

bench_row.py turns the output into a block to paste. It reads your local file, prints to the screen and sends nothing anywhere:

python3 bench_row.py run.json

Open a dataset Discussion and paste that block. Fill in the phone and network lines, saying whether airplane mode was on, and leave not recorded for anything you did not capture.

A run that failed to load or slowed down partway counts too. Say what happened.

Submitted results will be labelled with their model file and submitted_by, and kept separate from Neykosa's FTL measurements.

What a complete submission looks like

Here is our own measured run in the block format bench_row.py prints. We filled it in by hand from the source run above: that file is a summary, not raw llama-bench output, so the script did not print this. Yours will show your phone, the public model file and your own numbers.

Phone model, RAM, chip, Android version: Galaxy A05s SM-A057G, 4 GB, Snapdragon 680 'bengal', Android 15
Network state: not controlled (not airplane mode)
What went wrong, if anything: none
Model repository and revision: private Neykosa model, not published
Model file: qwen3-0.6b-v18-Q6_K.gguf
llama.cpp commit: 620d222
CPU as llama-bench reports it: not recorded
Settings: threads 4, type_k f16, type_v f16, flash_attn 0, repack 0

| Test | Depth | Mean tok/s | Standard deviation, tok/s |
| --- | --- | --- | --- |
| pp128 | d0 | 25.45 | 1.35 |
| tg64 | d0 | 14.82 | 0.16 |
| pp128 | d768 | 20.33 | 0.07 |
| tg64 | d768 | 11.02 | 0.01 |

How to read each line:

• Phone line. Give the exact model code from Settings > About phone: one retail name can cover different chips.
• Network state. Ours was not controlled, so we say so. Say whether airplane mode was on for yours.
• What went wrong. A crash, an out-of-memory stop or a slowdown partway is a result. Write it here.
• Model repository and revision. Yours names unsloth/Qwen3-0.6B-GGUF and the commit you downloaded. That is a different model from ours, so your row is labelled separately.
• CPU line. Our summary file did not keep llama-bench's CPU string, so it reads not recorded. Leave any line you lack the same way.
• Depth. Keep both depths: on our phone, writing falls from 14.82 tok/s at d0 to 11.02 tok/s at d768.
• Standard deviation. The spread over three repetitions. Our widest is pp128 at d0, sd 1.35. If yours is much wider, say what else was running.

Licences and notices

• Dataset: CC BY 4.0.
• Qwen3: Apache-2.0, by the Qwen team. The GGUF file above is published by Unsloth under the same licence.
• llama.cpp, including llama-bench and llama-quantize: MIT.

Team Neykosa · neykosa.com