Google Tensor G2 · 8–12 GB

How fast is Gemma 4 E2B on the Pixel 7?

About 9.1 tokens per second, faster than reading. The first word arrives after 1.5 seconds.

9.1tokens per second

Faster than reading

First word in 1.5 s · a 300-word answer in about 45 s

What 9.1 tokens per second looks like

Typical range 8.2–9.4 tokens per second · 10 phones

The same for the Pixel 7 Pro, with the same Google Tensor G2 chip.

Gemma 4 E2B on other Android phones

  1. Galaxy S25 UltraFirst word 0.4 s · 82 phones24tok/s
  2. Galaxy S26 UltraFirst word 0.6 s · 45 phones19tok/s
  3. Pixel 10 Pro XLFirst word 0.8 s · 23 phones16tok/s
  4. Galaxy S25 FEFirst word 1.0 s · 22 phones16tok/s
  5. Galaxy S24 UltraFirst word 1.2 s · 78 phones15tok/s
  6. Galaxy S23 UltraFirst word 21.8 s · 31 phones14tok/s
  7. Galaxy S22 UltraFirst word 2.4 s · 27 phones12tok/s
  8. Pixel 9First word 1.2 s · 32 phones11tok/s
  9. Galaxy S21 FE 5GFirst word 7.1 s · 19 phones11tok/s
  10. Pixel 8First word 1.2 s · 16 phones11tok/s
  11. TCL NXTPAPER 70 ProFirst word 2.5 s · 26 phones10tok/s
  12. Galaxy S20 FE 5GFirst word 1.6 s · 10 phones10tok/s
  13. moto g stylus 5GFirst word 1.7 s · 17 phones9.9tok/s
  14. Galaxy Note10+First word 2.3 s · 11 phones9.7tok/s
  15. moto g stylusFirst word 2.6 s · 25 phones9.4tok/s
  16. Pixel 7First word 1.5 s · 10 phones9.1tok/s
  17. moto g power 5GFirst word 3.5 s · 13 phones6.1tok/s
  18. moto g powerFirst word 5.3 s · 19 phones5.0tok/s

Run Gemma 4 E2B on your phone

Loci is free, needs no account, and keeps every conversation on your device.

Get it on Android

Last 30 days, through · updated daily

How these speeds are measured

  • Every number comes from real chats in Loci, timed on the device that wrote the answer, with Loci's default settings.
  • Tokens per second is how fast an answer is written once it starts. Four characters of text count as one token, for every model and device. Reading pace is about 5.
  • First word is how long an answer takes to start after you send a message. Longer conversations take longer.
  • Each device counts once. The large number is the median; the typical range covers the middle half of devices.
  • Answers written in Low Power Mode or Battery Saver, in the background, or with web search or other tools are left out.