family member
Llama 3.2 logo

Llama 3.2 1B

113 tokens a second on Mac · no GPU · 10 CPUs · 16 GB RAM. No hosted API sells it, so this network is the place to run it.

9 builds of Llama 3.2 1B, 35 signed runs on real machines: different packagers, precisions and engines, each a different file with its own numbers. 1.2B parameters, as the vendor states.

113tok/s, fastest machine
$0.81per M tokens · no API sells it
32Kcontext proved by recall
0.7 GBmemory, smallest build
8providers measured it
best on ollama · 9 builds

proved 32K context, by ollama Q8_0

113.3 tok/s peak, by maziyarpanahi Q4_K_M

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/sMac · no GPU · 10 CPUs · 16 GB RAM · Q4_K_M (4-bit) by maziyarpanahi · 113.27 tok/s · $0.8056/MMac · Apple silicon · 8 CPUs · 8 GB RAM · llama3.2:1b · 20.37 tok/s · $8.5986/MWindows · no GPU · 8 CPUs · 16 GB RAM · IQ3_XS (3-bit) by maziyarpanahi · 17.72 tok/s · $4.0054/MLinux · no GPU · 4 CPUs · 31 GB RAM · Q4_K_M (4-bit) by bartowski · 12.24 tok/s · $2.5368/MWindows · no GPU · 8 CPUs · 16 GB RAM · Q4_K_M (4-bit) by bartowski · 11.62 tok/s · $6.161/MLinux · no GPU · 4 CPUs · 31 GB RAM · llama3.2:1b · 11.45 tok/s · $2.6566/MLinux · no GPU · 4 CPUs · 31 GB RAM · Q4_K_M (4-bit) by hugging-quants · 11.25 tok/s · $2.7606/MLinux · no GPU · 4 CPUs · 31 GB RAM · llama3.2:1b · 10.79 tok/s · $2.8196/MLinux · no GPU · 4 CPUs · 31 GB RAM · Q3_K_L (3-bit) by bartowski · 10.3 tok/s · $3.007/MLinux · no GPU · 4 CPUs · 31 GB RAM · Q8_0 (8-bit) by bartowski · 8.46 tok/s · $3.717/MLinux · no GPU · 4 CPUs · 31 GB RAM · IQ1_S (1-bit) by maziyarpanahi · 6.6 tok/s · $4.6098/MWindows · AMD Radeon(TM) Vega 8 Mobile Graphics 1 GB · 8 CPUs · 7 GB RAM · llama3.2:1b · 5.62 tok/s · $31.1559/MLinux · no GPU · 4 CPUs · 31 GB RAM · IQ3_M (3-bit) by bartowski · 5.28 tok/s · $5.862/MMac · no GPU · 4 CPUs · 8 GB RAM · Q4_K_M (4-bit) by bartowski · 4.73 tok/s · $6.5625/MLinux · no GPU · 4 CPUs · 31 GB RAM · F16 (16-bit, full precision) by bartowski · 4.3 tok/s · $7.5285/MMac · Apple silicon · 4 CPUs · 8 GB RAM · llama3.2:1b · 3.37 tok/s · $39.9708/MMac · no GPU · 4 CPUs · 8 GB RAM · IQ1_S (1-bit) by maziyarpanahi · 3.26 tok/s · $9.3281/MMac · no GPU · 4 CPUs · 8 GB RAM · IQ3_XS (3-bit) by maziyarpanahi · 1.48 tok/s · $20.6089/M

Same build, different machines

Q8_0 by ollama on 17 machines: the fastest decodes 7.6× as fast as the slowest. One dot per machine, tokens per second.

  1. Mac, CPU only · 8 CPUs×229 tok/s
  2. Intel(R) Iris(R) Xe Graphics×113 tok/s
  3. Linux, CPU only · 4 CPUs×812 tok/s
  4. 2× Tesla T4×111 tok/s
  5. Intel(R) UHD Graphics 730×19.7 tok/s
  6. Linux, CPU only · 5 CPUs×18.6 tok/s
  7. AMD Radeon(TM) Vega 8 Mobile Graphics×18.3 tok/s
  8. Apple M4×16.2 tok/s
  9. Apple silicon×13.8 tok/s

What it fits on

The smallest build needs about 0.7 GB; 17 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 RAM1 of 7
  2. NVIDIA GeForce RTX 3050 Laptop GPU4 GB on the cardevery build
  3. NVIDIA GeForce GTX 16504 GB on the cardevery build
  4. AMD Radeon(TM) Vega 8 Mobile Graphics7 GB RAMevery build
  5. Windows, CPU only · 12 CPUs7 GB RAMevery build
  6. Apple silicon8 GB RAMevery build
  7. Mac, CPU only · 8 CPUs8 GB RAMevery build
  8. NVIDIA GeForce RTX 3080 Ti12 GB on the cardevery build
  9. Tesla T415 GB on the cardevery build
  10. Intel(R) UHD Graphics 73016 GB RAMevery build
  11. Intel(R) Iris(R) Xe Graphics16 GB RAMevery build
  12. Apple M416 GB RAMevery build
  13. Apple M116 GB RAMevery build
  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. 10 walks ended because the model forgot, 4 because the machine could not hold the prompt.

machine · build4K8K16K32K64K128Kwhy it stopped
Gram ·0548a8Mac, CPU only · 8 CPUs · Q8_0 by ollama32K an older probe: where it stopped was not recorded
nitin singhApple M4 · Q4_K_M by maziyarpanahi8K an older probe: where it stopped was not recorded
nitin singhApple M4 · Q4_K_M by bartowski8K an older probe: where it stopped was not recorded
Amit KhetanLinux, CPU only · 4 CPUs · Q8_0 by ollama8K too slow to answer here
nitin singhApple M4 · Q8_0 by ollama8K an older probe: where it stopped was not recorded
nitin singhApple M4 · IQ3_XS by maziyarpanahi4K an older probe: where it stopped was not recorded
Amit Khetan2× Tesla T4 · Q4_K_M by bartowski4K too slow to answer here at 8K
Amit KhetanIntel(R) UHD Graphics 730 · Q4_K_M by bartowski4K an older probe: where it stopped was not recorded
Amit Khetan2× Tesla T4 · Q8_0 by bartowski4K too slow to answer here
Gram ·d061c9AMD Radeon(TM) Vega 8 Mobile Graphics · Q8_0 by ollama4K too slow to answer here
nitin singhApple M4 · IQ1_S by maziyarpanahi— an older probe: where it stopped was not recorded
Amit KhetanIntel(R) UHD Graphics 730 · IQ3_XS by maziyarpanahi— an older probe: where it stopped was not recorded
Keshav AgarwalIntel(R) Iris(R) Xe Graphics · Q8_0 by ollama— an older probe: where it stopped was not recorded
Amit KhetanLinux, CPU only · 4 CPUs · Q8_0 by ollama— the model lost the needle at 4K

The 14 longest of 30 walks; every one is on its build's page.

Carbon per million tokens

Lowest build on each machine · 0 of 9 read off a power sensor, the rest modelled from the parts · fewer grams is better

  1. nitin singh85 g0.43 J/tok · modelled
  2. Amit Khetan189 g3.19 J/tok · modelled
  3. Amit Khetan191 g3.23 J/tok · modelled
  4. Amit Khetan200 g3.39 J/tok · modelled
  5. Amit Khetan212 g3.58 J/tok · modelled
  6. Gram ·73d610438 g3.33 J/tok · modelled
  7. Amit Khetan477 g3.63 J/tok · modelled
  8. Amit Khetan1.1 kg5.58 J/tok · modelled
  9. Gram ·d061c93.4 kg17.52 J/tok · modelled

Carbon is location-based: signed energy × the yearly average of the grid where the machine signed its place (India grid average · EU grid average (applied to NL) · EU grid average (applied to BE) · world grid average (Australia not in the table) · world grid average (place not signed); 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 · Structured output · Sustained output · Writes 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
Q8_0 by ollamallama3.2:1bollamaollama2.9 GB32K49.917
Q4_K_M by bartowskihf.co/bartowski/Llama-3.2-1B-Instruct-GGUF:Q4_K_Mollamaquantizer · bartowski1.1 GB16K103.94
Q4_K_M by maziyarpanahihf.co/MaziyarPanahi/Llama-3.2-1B-Instruct-GGUF:Q4_K_Mollamamaziyarpanahi—8K113.32
IQ3_XS by maziyarpanahihf.co/MaziyarPanahi/Llama-3.2-1B-Instruct-GGUF:IQ3_XSollamamaziyarpanahi—4K107.53
Q8_0 by bartowskihf.co/bartowski/Llama-3.2-1B-Instruct-GGUF:Q8_0ollamaquantizer · bartowski1.7 GB4K8.71
IQ1_S by maziyarpanahihf.co/MaziyarPanahi/Llama-3.2-1B-Instruct-GGUF:IQ1_Sollamamaziyarpanahi0.7 GB—105.04
Q3_K_L by bartowskihf.co/bartowski/Llama-3.2-1B-Instruct-GGUF:Q3_K_Lollamaquantizer · bartowski1.0 GB—10.41
F16 by bartowskihf.co/bartowski/Llama-3.2-1B-Instruct-GGUF:F16ollamaquantizer · bartowski3.0 GB—8.02
IQ3_M by bartowskihf.co/bartowski/Llama-3.2-1B-Instruct-GGUF:IQ3_Mollamaquantizer · bartowski1.0 GB—5.31
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.