Gemma 4 E2B
291 tokens a second on Linux · RTX PRO 6000 Blackwell Server Edition 96 GB · 48 CPUs · 177 GB RAM. No hosted API sells it, so this network is the place to run it.
13 builds of Gemma 4 E2B, 55 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.
Speed against price, every machine
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
- Tesla T4×178 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; 17 of 18 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 7
- NVIDIA GeForce GTX 16504 GB on the card3 of 7
- AMD Radeon(TM) Vega 8 Mobile Graphics7 GB RAM6 of 7
- Windows, CPU only · 12 CPUs7 GB RAM6 of 7
- Apple silicon8 GB RAM6 of 7
- Apple M18 GB RAM6 of 7
- Mac, CPU only · 8 CPUs8 GB RAM6 of 7
- NVIDIA GeForce RTX 3080 Ti12 GB on the card6 of 7
- Tesla T415 GB on the card6 of 7
- Intel(R) UHD Graphics 73016 GB RAM6 of 7
- Intel(R) Iris(R) Xe Graphics16 GB RAM6 of 7
- Apple M416 GB RAM6 of 7
- Apple M116 GB RAM6 of 7
- 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, 10 because the machine could not hold the prompt.
The 14 longest of 52 walks; every one is on its build's page.
Carbon per million tokens
Lowest build on each machine · 5 of 6 read off a power sensor, the rest modelled from the parts · fewer grams is better
- Gram ·dae67996 g0.73 J/tok · sensor
- Gram ·e189d2132 g2.23 J/tok · sensor
- Gram ·a5be46133 g1.24 J/tok · sensor
- Gram ·e61d3b257 g4.34 J/tok · sensor
- Gram ·c76d30544 g4.14 J/tok · sensor
- Gram ·afbb8a740 g3.76 J/tok · modelled
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.
Measured by
The people whose machines signed these numbers
- Gram ·9e3c799 runs
- Gram ·a5be467 runs
- Gram ·e61d3b7 runs
- Gram ·afbb8a6 runs
- Gram ·84bf544 runs
- Gram ·d686724 runs
- Gram ·dae6793 runs
- Gram ·e189d22 runs
- Gram ·e78fd22 runs
- Gram ·c76d302 runs
and 6 more
proved by at least one build: Tool calling · Tool loop finished · Faithful to tool results · Chained tool calls · Structured output · Sustained output · Extended thinking · Vision · Counts in images · Multiple images at once · Spatial reasoning · Writes code · Reasons about code · Synthesises sources · Follows instructions
Every build
Each its own file with its own page: the machines that ran it and their signatures.
| build | engine | whose work | memory | context proved | best tok/s | machines |
|---|---|---|---|---|---|---|
| Q4_K_M by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:q4_k_m | ollama | quantizer · unsloth | 4.9 GB | 128K | 291.2 | 13 |
| Q4_K_M by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:Q4_K_M | ollama | quantizer · unsloth | 4.9 GB | 128K | 291.2 | 13 |
| IQ4_XS by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:IQ4_XS | ollama | quantizer · unsloth | 3.9 GB | 128K | 182.9 | 3 |
| Q4_K_M by googlegemma4:e2b | ollama | vendor · google | 1.4 GB | 128K | 170.5 | 10 |
| Q4_0 by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:Q4_0 | ollama | quantizer · unsloth | 2.9 GB | 128K | 84.7 | 3 |
| Q8_0 by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:Q8_0 | ollama | quantizer · unsloth | 6.4 GB | 128K | 67.0 | 2 |
| BF16 by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:BF16 | ollama | quantizer · unsloth | 16.7 GB | 128K | 38.3 | 2 |
| IQ4_NL by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:IQ4_NL | ollama | quantizer · unsloth | — | 8K | 54.9 | 2 |
| Q3_K_M by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:Q3_K_M | ollama | quantizer · unsloth | — | 8K | 38.4 | 3 |
| Q3_K_S by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:Q3_K_S | ollama | quantizer · unsloth | — | 4K | 14.4 | 1 |
| by ollamagemma4:e2b-it-qat | ollama | ollama | — | — | 8.9 | 1 |
| by googlegoogle/gemma-4-E2B-it-qat-q4_0-gguf | llamacpp | — | — | 2.0 | 1 | |
| Q2_K_XL by unslothunsloth/gemma-4-E2B-it-qat-mobile-GGUF:Q2_K_XL | llamacpp | unsloth | — | — | — | 1 |
Fine-tunes, merges and abliterations of this member, each its own model with its own page. None of their numbers are counted above.