291 tokens a second on a Linux PC with an NVIDIA GPU, 256 GB. No hosted API sells it, so this network is the place to run it.
12 builds of Gemma 4 E2B, 42 signed runs on real machines: different packagers, precisions and engines, each a different file with its own numbers. 5.1B parameters, as the vendor states.
Up and left is faster and cheaper · the line joins machines no other beats on both · price on a log scale
Same build, different machines
Q4_K_M by unsloth on 13 machines: the fastest decodes 73× as fast as the slowest. One dot per machine, tokens per second.
NVIDIA RTX PRO 6000 Blackwell Server Edition×1291 tok/s
NVIDIA L4×1106 tok/s
NVIDIA GeForce RTX 3080 Ti×193 tok/s
2× Tesla T4×380 tok/s
Linux, CPU only · 4 CPUs×178 tok/s
Apple M4×155 tok/s
NVIDIA GeForce RTX 3050 Laptop GPU×139 tok/s
Windows, CPU only · 12 CPUs×115 tok/s
NVIDIA GeForce GTX 1650×111 tok/s
Intel(R) UHD Graphics 730×16.8 tok/s
Apple silicon×14.0 tok/s
What it fits on
The smallest build needs about 1.4 GB; 16 of 17 kinds of machine on the network can hold it. Bars: each kind's memory (its card's, else RAM) · ticks: every build, at the size it held plus a fifth for the prompt.
Linux, CPU only · 2 CPUs1 GB RAMnone
NVIDIA GeForce RTX 3050 Laptop GPU4 GB on the card3 of 6
NVIDIA GeForce GTX 16504 GB on the card3 of 6
AMD Radeon(TM) Vega 8 Mobile Graphics7 GB RAM5 of 6
Windows, CPU only · 12 CPUs7 GB RAM5 of 6
Apple silicon8 GB RAM5 of 6
Mac, CPU only · 8 CPUs8 GB RAM5 of 6
NVIDIA GeForce RTX 3080 Ti12 GB on the card5 of 6
Tesla T415 GB on the card5 of 6
Intel(R) UHD Graphics 73016 GB RAM5 of 6
Intel(R) Iris(R) Xe Graphics16 GB RAM5 of 6
Apple M416 GB RAM5 of 6
Apple M116 GB RAM5 of 6
NVIDIA L422 GB on the cardevery build
2× Tesla T429 GB on the cardevery build
Linux, CPU only · 4 CPUs31 GB RAMevery build
NVIDIA RTX PRO 6000 Blackwell Server Edition95 GB on the cardevery build
How far back it remembers, machine by machine
Filled: the needle came back at that prompt size · outlined: asked and lost · blank: not asked · trained for 128K. 1 walk ended because the model forgot, 7 because the machine could not hold the prompt.
Carbon is location-based: signed energy × the yearly average of the grid where the machine signed its place (world grid average (Singapore not in the table) · EU grid average (applied to NL) · United States grid average · world grid average (TW not in the table) · India grid average; Ember Global Electricity Review 2025 (2024 generation), location-based). Facility overhead is not included. A machine on its own solar is not credited until it can prove it.