family member
Qwen 3 logo

Qwen 3 4B

90 tokens a second on Windows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM. No hosted API sells it, so this network is the place to run it.

13 builds of Qwen 3 4B, 29 signed runs on real machines: different packagers, precisions and engines, each a different file with its own numbers.

141tok/s, fastest machine
$2.21per M tokens · no API sells it
64Kcontext proved by recall
2.0 GBmemory, smallest build
7providers measured it
Run it on your gram Install a gram with one command. It works through the catalog on your own machine and signs every result. 7 providers have measured Qwen 3 4B so far.Use it in your own room Ask Evo in your gramx and it runs on a machine that proved it, at the metered $2.21 per million tokens. On your own machines it costs you nothing.Compare every build Put it beside every other model in one table: speed, price, context and the API price, one row each.
best on ollama · 12 builds

trained 40K · proved 40K context, by Qwen Q4_K_M

140.7 tok/s peak, by Qwen Q4_K_M

best on vllm · 1 build

proved 64K context, by qwen

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

$10USD per million tokenstok/sWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · 1-bit by unsloth · 90.41 tok/s · $2.8436/MWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · 16-bit (full precision) by unsloth · 80.76 tok/s · $3.2502/MLinux · NVIDIA 15 GB · 4 CPUs · 31 GB RAM · 1-bit by unsloth · 61.33 tok/s · $2.2082/MWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · 3-bit by unsloth · 53.3 tok/s · $4.8319/MLinux · NVIDIA 15 GB · 4 CPUs · 31 GB RAM · 4-bit by Qwen · 50.44 tok/s · $2.7067/MLinux · L4 23 GB · 12 CPUs · 53 GB RAM · 4-bit by Qwen · 49.84 tok/s · $4.3667/MLinux · NVIDIA 15 GB · 4 CPUs · 31 GB RAM · 5-bit by Qwen · 47.96 tok/s · $2.8516/MLinux · NVIDIA 15 GB · 4 CPUs · 31 GB RAM · 3-bit by unsloth · 47.81 tok/s · $2.8422/MLinux · NVIDIA 15 GB · 4 CPUs · 31 GB RAM · 5-bit by Qwen · 46.99 tok/s · $2.9117/MLinux · NVIDIA 15 GB · 4 CPUs · 31 GB RAM · 8-bit by Qwen · 38.86 tok/s · $3.5489/MWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · 5-bit by Qwen · 35.37 tok/s · $7.3075/MMac · Apple silicon · 10 CPUs · 16 GB RAM · 5-bit by Qwen · 22.66 tok/s · $8.7188/MWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · 8-bit by Qwen · 17.06 tok/s · $15.2173/MMac · Apple silicon · 10 CPUs · 16 GB RAM · 1-bit by unsloth · 14.37 tok/s · $13.6614/MWindows · NVIDIA 12 GB · 16 CPUs · 16 GB RAM · 5-bit by Qwen · 11.16 tok/s · $23.165/MWindows · Intel 2 GB · 8 CPUs · 16 GB RAM · 3-bit by unsloth · 4.63 tok/s · $15.5931/M

Same build, different machines

IQ3_XXS by unsloth on 4 machines: the fastest decodes 286× as fast as the slowest. One dot per machine, tokens per second.

  1. NVIDIA GeForce RTX 3080 Ti×163 tok/s
  2. 2× Tesla T4×149 tok/s
  3. Intel(R) UHD Graphics 730×14.6 tok/s
  4. Apple M4×10.2 tok/s

What it fits on

The smallest build needs about 2.0 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.

  1. Linux, CPU only · 2 CPUs1 GB RAMnone
  2. NVIDIA GeForce RTX 3050 Laptop GPU4 GB on the card1 of 7
  3. NVIDIA GeForce GTX 16504 GB on the card1 of 7
  4. AMD Radeon(TM) Vega 8 Mobile Graphics7 GB RAM1 of 7
  5. Windows, CPU only · 12 CPUs7 GB RAM1 of 7
  6. Apple silicon8 GB RAM1 of 7
  7. Apple M18 GB RAM1 of 7
  8. Mac, CPU only · 8 CPUs8 GB RAM1 of 7
  9. NVIDIA GeForce RTX 3080 Ti12 GB on the card5 of 7
  10. Tesla T415 GB on the cardevery build
  11. Intel(R) UHD Graphics 73016 GB RAMevery build
  12. Intel(R) Iris(R) Xe Graphics16 GB RAMevery build
  13. Apple M416 GB RAMevery build
  14. Apple M116 GB RAMevery build
  15. NVIDIA L422 GB on the cardevery build
  16. 2× Tesla T429 GB on the cardevery build
  17. Linux, CPU only · 4 CPUs31 GB RAMevery build
  18. 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 40K. 2 walks ended because the model forgot, 1 because the machine could not hold the prompt.

machine · build4K8K16K32K40K64K128Kwhy it stopped
Gram ·a5be462× Tesla T4 · by qwen64K this machine could not serve at 128K
Gram ·dae679NVIDIA L4 · Q4_K_M by Qwen40K recalled at every size asked
Gram ·a5be462× Tesla T4 · IQ3_XXS by unsloth32K recalled at every size asked
Gram ·a5be462× Tesla T4 · Q5_0 by Qwen32K recalled at every size asked
Gram ·a5be462× Tesla T4 · Q5_K_M by Qwen32K recalled at every size asked
Gram ·a5be462× Tesla T4 · Q8_0 by Qwen32K recalled at every size asked
Gram ·84bf54NVIDIA GeForce RTX 3080 Ti · BF16 by unsloth16K an older probe: where it stopped was not recorded
Gram ·a5be462× Tesla T4 · Q4_K_M by Qwen16K the model lost the needle at 32K
Gram ·84bf54NVIDIA GeForce RTX 3080 Ti · IQ3_XXS by unsloth4K an older probe: where it stopped was not recorded
Gram ·afbb8aApple M4 · Q4_0 by unsloth4K an older probe: where it stopped was not recorded
Gram ·afbb8aApple M4 · Q4_1 by unsloth4K an older probe: where it stopped was not recorded
Gram ·afbb8aApple M4 · Q5_0 by Qwen4K an older probe: where it stopped was not recorded
Gram ·9e3c79Intel(R) UHD Graphics 730 · Q4_0 by unsloth4K an older probe: where it stopped was not recorded
Gram ·9e3c79Intel(R) UHD Graphics 730 · IQ4_XS by unsloth4K an older probe: where it stopped was not recorded

The 14 longest of 25 walks; every one is on its build's page.

Carbon per million tokens

Lowest build on each machine · 2 of 2 read off a power sensor, the rest modelled from the parts · fewer grams is better

  1. Gram ·a5be46211 g1.98 J/tok · sensor
  2. Gram ·dae679308 g2.34 J/tok · sensor

Carbon is location-based: signed energy × the yearly average of the grid where the machine signed its place (United States grid average · world grid average (Singapore not in the table); 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

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 · 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
by qwenQwen/Qwen3-4Bvllmqwen—64K—1
Q4_K_M by Qwenhf.co/Qwen/Qwen3-4B-GGUF:Q4_K_Mollamavendor · Qwen10.1 GB40K140.73
IQ3_XXS by unslothhf.co/unsloth/Qwen3-4B-GGUF:IQ3_XXSollamaquantizer · unsloth13.1 GB32K62.94
Q5_0 by Qwenhf.co/Qwen/Qwen3-4B-GGUF:Q5_0ollamavendor · Qwen8.8 GB32K48.53
Q5_K_M by Qwenhf.co/Qwen/Qwen3-4B-GGUF:Q5_K_Mollamavendor · Qwen8.9 GB32K47.02
Q8_0 by Qwenhf.co/Qwen/Qwen3-4B-GGUF:Q8_0ollamavendor · Qwen10.4 GB32K38.92
BF16 by unslothhf.co/unsloth/Qwen3-4B-GGUF:BF16ollamaquantizer · unsloth12.6 GB16K80.81
Q4_0 by unslothhf.co/unsloth/Qwen3-VL-4B-Instruct-GGUF:Q4_0ollamaunsloth—4K40.94
Q4_1 by unslothhf.co/unsloth/Qwen3-VL-4B-Instruct-GGUF:Q4_1ollamaunsloth—4K37.02
IQ4_XS by unslothhf.co/unsloth/Qwen3-VL-4B-Instruct-GGUF:IQ4_XSollamaunsloth—4K6.41
IQ1_S by unslothhf.co/unsloth/Qwen3-4B-GGUF:IQ1_Sollamaquantizer · unsloth2.0 GB—90.43
Q4_1 by unslothhf.co/unsloth/Qwen3-4B-GGUF:Q4_1ollamaquantizer · unsloth——19.42
Q2_K by unslothhf.co/unsloth/Qwen3-VL-4B-Instruct-GGUF:Q2_Kollamaunsloth——11.91