Benchmark collection for different AI models and devices

[aryoda]

I want to collect reported benchmark results in this "issue" as a table.

Please feel free to contribute by publishing your results as answers so that I can add it to this table...

Please do not close this issue until we have enough results...

Device name	OS Version	Box version	Gemma-4-E4B-it (on GPU)	Gemma-4-E4B-it (on NPU)	Gemma-4-E4B-it (on CPU)	Notes
Google Pixel 9 Pro XL	GrapheneOS build 2026042101 (Android 16)	v1.0.3custom	Prefill: 120 TPS Decode: 9,54 TPS Time to first token: 2,23 s	Prefill: 19,31 TPS Decode: 8,11 TPS Time to first token: 13,40 s	Prefill: 16,74 TPS Decode: 7,40 TPS Time to first token: 15,89 s	Inference on CPU causes the device to become hot (not observed with GPU and NPU)

Legend:
TPS = tokens/second

Note: This table is maintained via tablesgenerator

[jegly]

Device: Google Pixel 10 Pro Fold | OS: GrapheneOS | App: v1.0.3 custom-rom-support | Model: Gemma-4-E2B-it

┌─────────────────┬─────────────┬─────────────┬───────┐
│ Accelerator │ Prefill │ Decode │ TTFT │
├─────────────────┼─────────────┼─────────────┼───────┤
│ TPU (via NNAPI) │ 51.38 tok/s │ 17.18 tok/s │ 5.09s │
└─────────────────┴─────────────┴─────────────┴───────┘

Config: 256 prefill/decode tokens, 3 runs

---

Device: Google Pixel 10 Pro Fold | OS: Stock Android 17 | App: v1.0.4 main | Model: Gemma-4-E2B-it

┌─────────────┬──────────────┬─────────────┬───────┬─────────────┐
│ Accelerator │ Prefill │ Decode │ TTFT │ Steady Init │
├─────────────┼──────────────┼─────────────┼───────┼─────────────┤
│ GPU │ 259.71 tok/s │ 8.39 tok/s │ 1.11s │ 3.07s │
├─────────────┼──────────────┼─────────────┼───────┼─────────────┤
│ CPU │ 54.35 tok/s │ 16.33 tok/s │ 4.73s │ 1.01s │
└─────────────┴──────────────┴─────────────┴───────┴─────────────┘

Config: 256 prefill/decode tokens, 3 runs

---

Device: Google Pixel 6a | OS: Stock Android 17 | App: v1.0.4 main | Model: Gemma-4-E2B-it

┌─────────────┬──────────────┬────────────┬───────┬─────────────┐
│ Accelerator │ Prefill │ Decode │ TTFT │ Steady Init │
├─────────────┼──────────────┼────────────┼───────┼─────────────┤
│ GPU │ 219.29 tok/s │ 9.45 tok/s │ 1.28s │ 6.09s │
└─────────────┴──────────────┴────────────┴───────┴─────────────┘

Config: 256 prefill/decode tokens, 3 runs