family member
Gemma 4 logo

Gemma 4 E2B

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.

291tok/s, fastest machine
$0.43per M tokens · no API sells it
128Kcontext proved by recall
1.4 GBmemory, smallest build
12providers measured it
best on ollama · 10 builds

proved 128K context, by unsloth Q4_K_M

291.2 tok/s peak, by unsloth Q4_K_M

best on llamacpp · 2 builds

no context proved on this engine yet

2.0 tok/s peak, by google

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

$1$10USD per million tokenstok/sLinux PC with an NVIDIA GPU, 256 GB · Q4_K_M (4-bit) by unsloth · 291.17 tok/s · $1.6475/MLinux · RTX PRO 6000 Blackwell Server Edition 96 GB · 48 CPUs · 177 GB RAM · Q4_K_M (4-bit) by unsloth · 291.17 tok/s · $2.0053/MWindows · GeForce RTX 3080 Ti 12 GB · 16 CPUs · 16 GB RAM · IQ4_XS (4-bit) by unsloth · 182.88 tok/s · $1.4139/MWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · Q4_K_M (4-bit) by google · 170.45 tok/s · $1.5375/MLinux · L4 23 GB · 12 CPUs · 53 GB RAM · Q4_K_M (4-bit) by google · 149.89 tok/s · $1.4779/MLinux PC with an NVIDIA GPU, 64 GB · Q4_K_M (4-bit) by unsloth · 98.85 tok/s · $1.1604/MLinux · L4 23 GB · 12 CPUs · 53 GB RAM · Q4_K_M (4-bit) by unsloth · 98.85 tok/s · $2.2141/MWindows PC with an NVIDIA GPU, 16 GB · Q4_K_M (4-bit) by unsloth · 97.95 tok/s · $2.6485/MLinux · 2× Tesla T4 15 GB · 4 CPUs · 31 GB RAM · Q4_K_M (4-bit) by google · 91.08 tok/s · $1.5416/MWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · Q4_K_M (4-bit) by unsloth · 87.36 tok/s · $2.9697/MLinux · NVIDIA 15 GB · 4 CPUs · 31 GB RAM · Q4_0 (4-bit) by unsloth · 83.82 tok/s · $1.6269/MLinux · 2× Tesla T4 15 GB · 4 CPUs · 31 GB RAM · IQ4_XS (4-bit) by unsloth · 78.69 tok/s · $1.7397/MLinux PC with an NVIDIA GPU, 32 GB · Q4_K_M (4-bit) by unsloth · 77.58 tok/s · $0.433/MLinux · NVIDIA 15 GB · 4 CPUs · 31 GB RAM · Q4_K_M (4-bit) by unsloth · 77.58 tok/s · $1.7757/MLinux · 2× Tesla T4 15 GB · 4 CPUs · 31 GB RAM · Q4_K_M (4-bit) by unsloth · 74.18 tok/s · $1.8572/MMac · no GPU · 10 CPUs · 16 GB RAM · Q4_0 (4-bit) by unsloth · 57.55 tok/s · $1.6164/MMac with Apple silicon, 16 GB · Q4_K_M (4-bit) by unsloth · 53.69 tok/s · $1.381/MMac · Apple M4 · 10 CPUs · 16 GB RAM · Q4_K_M (4-bit) by unsloth · 53.69 tok/s · $3.6987/MWindows · NVIDIA 4 GB · 12 CPUs · 16 GB RAM · Q4_K_M (4-bit) by unsloth · 28.05 tok/s · $7.8028/MLinux PC with no GPU, 32 GB · Q4_K_M (4-bit) by unsloth · 24.89 tok/s · $1.3496/MLinux · no GPU · 4 CPUs · 31 GB RAM · Q4_K_M (4-bit) by unsloth · 24.89 tok/s · $1.3495/MWindows PC with an NVIDIA GPU, 8 GB · Q4_K_M (4-bit) by unsloth · 15.33 tok/s · $9.6476/MWindows · NVIDIA 4 GB · 12 CPUs · 7 GB RAM · Q4_K_M (4-bit) by unsloth · 15.33 tok/s · $14.2764/MWindows · Intel 2 GB · 8 CPUs · 16 GB RAM · Q3_K_S (3-bit) by unsloth · 13 tok/s · $5.5982/MWindows · NVIDIA 4 GB · 8 CPUs · 8 GB RAM · Q4_K_M (4-bit) by unsloth · 9.95 tok/s · $17.9178/MWindows · no GPU · 8 CPUs · 16 GB RAM · Q4_0 (4-bit) by unsloth · 9.1 tok/s · $7.9949/MLinux · no GPU · 4 CPUs · 31 GB RAM · Q4_K_M (4-bit) by google · 7.15 tok/s · $5.0695/MLinux · no GPU · 4 CPUs · 31 GB RAM · IQ4_XS (4-bit) by unsloth · 6.93 tok/s · $4.7232/MWindows PC with Intel graphics, 16 GB · Q4_K_M (4-bit) by unsloth · 6.8 tok/s · $10.9041/MWindows · Intel 2 GB · 8 CPUs · 16 GB RAM · Q4_K_M (4-bit) by unsloth · 6.8 tok/s · $10.9087/MLinux · no GPU · 4 CPUs · 31 GB RAM · Q8_0 (8-bit) by unsloth · 5.33 tok/s · $6.4387/MIntel Mac, 8 GB · Q4_K_M (4-bit) by unsloth · 3.99 tok/s · $8.4191/MLinux · no GPU · 4 CPUs · 31 GB RAM · BF16 (16-bit, full precision) by unsloth · 3.14 tok/s · $11.9727/MMac · no GPU · 4 CPUs · 8 GB RAM · Q3_K_M (3-bit) by unsloth · 2.78 tok/s · $11.5908/M

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.

  1. NVIDIA RTX PRO 6000 Blackwell Server Edition×1291 tok/s
  2. NVIDIA L4×1106 tok/s
  3. NVIDIA GeForce RTX 3080 Ti×193 tok/s
  4. 2× Tesla T4×380 tok/s
  5. Linux, CPU only · 4 CPUs×178 tok/s
  6. Apple M4×155 tok/s
  7. NVIDIA GeForce RTX 3050 Laptop GPU×139 tok/s
  8. Windows, CPU only · 12 CPUs×115 tok/s
  9. NVIDIA GeForce GTX 1650×111 tok/s
  10. Intel(R) UHD Graphics 730×16.8 tok/s
  11. 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.

  1. Linux, CPU only · 2 CPUs1 GB RAMnone
  2. NVIDIA GeForce RTX 3050 Laptop GPU4 GB on the card3 of 6
  3. NVIDIA GeForce GTX 16504 GB on the card3 of 6
  4. AMD Radeon(TM) Vega 8 Mobile Graphics7 GB RAM5 of 6
  5. Windows, CPU only · 12 CPUs7 GB RAM5 of 6
  6. Apple silicon8 GB RAM5 of 6
  7. Mac, CPU only · 8 CPUs8 GB RAM5 of 6
  8. NVIDIA GeForce RTX 3080 Ti12 GB on the card5 of 6
  9. Tesla T415 GB on the card5 of 6
  10. Intel(R) UHD Graphics 73016 GB RAM5 of 6
  11. Intel(R) Iris(R) Xe Graphics16 GB RAM5 of 6
  12. Apple M416 GB RAM5 of 6
  13. Apple M116 GB RAM5 of 6
  14. NVIDIA L422 GB on the cardevery build
  15. 2× Tesla T429 GB on the cardevery build
  16. Linux, CPU only · 4 CPUs31 GB RAMevery build
  17. 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.

machine · build4K8K16K32K64K128Kwhy it stopped
Gram ·e189d2NVIDIA RTX PRO 6000 Blackwell Server Edition · Q4_K_M by unsloth128K recalled at every size asked
Amit KhetanNVIDIA L4 · Q4_K_M by google128K recalled at every size asked
Amit Khetan2× Tesla T4 · Q4_K_M by google128K recalled at every size asked
Amit KhetanNVIDIA L4 · Q4_K_M by unsloth128K recalled at every size asked
Kunal HindochaNVIDIA GeForce RTX 3080 Ti · Q4_K_M by unsloth128K an older probe: where it stopped was not recorded
Amit Khetan2× Tesla T4 · Q4_0 by unsloth128K recalled at every size asked
Amit Khetan2× Tesla T4 · IQ4_XS by unsloth128K recalled at every size asked
Gram ·c76d302× Tesla T4 · Q4_K_M by unsloth128K recalled at every size asked
Amit Khetan2× Tesla T4 · Q4_K_M by unsloth128K recalled at every size asked
Amit Khetan2× Tesla T4 · Q8_0 by unsloth128K recalled at every size asked
Amit Khetan2× Tesla T4 · BF16 by unsloth128K recalled at every size asked
nitin singhApple M4 · Q4_K_M by google8K an older probe: where it stopped was not recorded
nitin singhApple M4 · Q3_K_M by unsloth8K an older probe: where it stopped was not recorded
Amit Khetan2× Tesla T4 · Q4_K_M by google8K too slow to answer here

The 14 longest of 39 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

  1. Amit Khetan96 g0.73 J/tok · sensor
  2. Gram ·e189d2132 g2.23 J/tok · sensor
  3. Amit Khetan133 g1.24 J/tok · sensor
  4. Amit Khetan257 g4.34 J/tok · sensor
  5. Gram ·c76d30544 g4.14 J/tok · sensor
  6. nitin singh740 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

and 2 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.

buildenginewhose workmemorycontext provedbest tok/smachines
Q4_K_M by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:q4_k_mollamaquantizer · unsloth4.9 GB128K291.213
IQ4_XS by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:IQ4_XSollamaquantizer · unsloth3.9 GB128K182.93
Q4_K_M by googlegemma4:e2bollamavendor · google1.4 GB128K170.510
Q4_0 by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:Q4_0ollamaquantizer · unsloth2.9 GB128K84.73
Q8_0 by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:Q8_0ollamaquantizer · unsloth6.4 GB128K67.02
BF16 by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:BF16ollamaquantizer · unsloth16.7 GB128K38.32
IQ4_NL by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:IQ4_NLollamaquantizer · unsloth—8K54.92
Q3_K_M by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:Q3_K_Mollamaquantizer · unsloth—8K38.43
Q3_K_S by unslothhf.co/unsloth/gemma-4-E2B-it-GGUF:Q3_K_Sollamaquantizer · unsloth—4K14.41
by ollamagemma4:e2b-it-qatollamaollama——8.91
by googlegoogle/gemma-4-E2B-it-qat-q4_0-ggufllamacppgoogle——2.01
Q2_K_XL by unslothunsloth/gemma-4-E2B-it-qat-mobile-GGUF:Q2_K_XLllamacppunsloth———1
derivatives — somebody else's change to these weights

Fine-tunes, merges and abliterations of this member, each its own model with its own page. None of their numbers are counted above.