On-device benchmark leaderboard
Every row is one configuration of the standard preset. Rows with more than one sample show a median and the sample count. Provenance is on every row — we do not blend measured and estimated numbers.
Not sure which model suits your phone? Start with your device.
Showing 14 of 14 configurations.
| Chip | Quant | Runtime | Ctx | Source | |||||
|---|---|---|---|---|---|---|---|---|---|
| samsung SM-S938B | Snapdragon 8 Elite | Gemma 4 E2B (Q4) | INT4 | litert/cpu | 4096 | 0.35s | 87.2 t/s | 1.2 GB | community |
| samsung SM-S938B | Snapdragon 8 Elite | MiniCPM5 1B (Q4) | Q4_K_M | gguf/cpu | 4096 | 0.20s | 59.1 t/s | 1.5 GB | community |
| Galaxy S24 Ultrasamsung SM-S928B | Snapdragon 8 Gen 3 | LFM2.5 1.2B Instruct (Q4) | INT4 | litert/cpu | 4096 | 2.17s | 59.1 t/s | 1.4 GB | community |
| Galaxy S24 Ultrasamsung SM-S928B | Snapdragon 8 Gen 3 | Gemma 4 E2B (Q4) | INT4 | litert/cpu | 4096 | 0.62s | 55.1 t/s | 1.1 GB | community |
| samsung SM-S938B | Snapdragon 8 Elite | Gemma 3 1B (Q4) | Q4_K_M | gguf/cpu | 4096 | 0.47s | 45.3 t/s | 1.7 GB | community |
| Galaxy S24 Ultrasamsung SM-S928B | Snapdragon 8 Gen 3 | MiniCPM5 1B (Q4) | Q4_K_M | gguf/cpu | 4096 | 0.32s | 36.9 t/s | 1.4 GB | community |
| Galaxy S21 Ultrasamsung SM-G998B | exynos2100 | Gemma 4 E2B (Q4) | INT4 | litert/cpu | 4096 | 0.90s | 29.3 t/s | 2.9 GB | community |
| Google Pixel 6 Pro | raven | Gemma 4 E2B (Q4) | INT4 | litert/cpu | 4096 | 0.79s | 27.8 t/s | 2.9 GB | community |
| Google Pixel 6 Pro | raven | MiniCPM5 1B (Q4) | Q4_K_M | gguf/cpu | 4096 | 0.70s | 25.4 t/sn=2 | 1.6 GB | community |
| Galaxy S24 Ultrasamsung SM-S928B | Snapdragon 8 Gen 3 | LFM2.5 2.6B (Q4_0 ARM) | Q4_0 | gguf/cpu | 4096 | 1.73s | 17.8 t/sn=2 | 3.4 GB | community |
| Galaxy S21 Ultrasamsung SM-G998B | exynos2100 | MiniCPM5 1B (Q4) | Q4_K_M | gguf/cpu | 4096 | 1.23s | 17.5 t/s | 1.6 GB | community |
| Google Pixel 6 Pro | raven | Qwen3.5 2B (Q4_0 ARM) | Q4_0 | gguf/cpu | 4096 | 6.93s | 12.7 t/s | 2.7 GB | community |
| Google Pixel 6 Pro | raven | LFM2.5 2.6B (Q4_0 ARM) | Q4_0 | gguf/cpu | 4096 | 11.15s | 11.7 t/s | 3.6 GB | community |
| Google Pixel 6 Pro | raven | SmolLM3 3B (Q4_0 ARM/GPU) | Q4_0 | gguf/cpu | 4096 | 3.86s | 8.3 t/s | 3.9 GB | community |
Custom-preset runs
These were measured on real hardware but not under the standard preset, so they are not comparable with the table above or with each other — a shorter generation is a faster number for reasons that have nothing to do with the device. They are listed because they are real measurements of real hardware, and kept separate because they cannot be ranked against anything. Nothing on this site derives a device recommendation from them.
| Device | Chip | Model | Preset | Quant | Runtime | Ctx | TTFT | Decode | Peak RAM | Source |
|---|---|---|---|---|---|---|---|---|---|---|
| Xiaomi 2506BPN68G | klimt | Phi-4 Mini (Q4) | parameters not recorded | Q8 | litert/cpu | 8192 | 6.01s | 10.3 t/s | 6.0 GB | community |