family member
gpt-oss logo

gpt-oss 20B

55 tokens a second on Linux · 2× Tesla T4 15 GB · 4 CPUs · 31 GB RAM. About $2.62 per million tokens here, against OpenRouter's $0.13: the API is 20× cheaper. Run it here for privacy or to keep work on your own machines, not to save money.

6 builds of gpt-oss 20B, 18 signed runs on real machines: different packagers, precisions and engines, each a different file with its own numbers.

73tok/s, fastest machine
$2.62per M tokens · $0.13 on an API
128Kcontext proved by recall
12.9 GBmemory, smallest build
8providers 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. 8 providers have measured gpt-oss 20B 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 $2.62 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 · 6 builds

trained 128K · proved 128K context, by unsloth Q4_K_M

73.2 tok/s peak, by unsloth 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 · dashed: OpenRouter

$0.1$1$10USD per million tokenstok/sLinux · 2× Tesla T4 15 GB · 4 CPUs · 31 GB RAM · 3-bit by unsloth · 54.73 tok/s · $2.6214/MLinux · 2× Tesla T4 15 GB · 4 CPUs · 31 GB RAM · 4-bit by unsloth · 52.39 tok/s · $2.7409/MLinux · 2× Tesla T4 15 GB · 4 CPUs · 31 GB RAM · 8-bit by unsloth · 44.16 tok/s · $3.2797/MLinux · L4 23 GB · 12 CPUs · 53 GB RAM · 4-bit by unsloth · 24.03 tok/s · $9.3514/MLinux · no GPU · 4 CPUs · 31 GB RAM · 4-bit by unsloth · 4.56 tok/s · $8.6295/MLinux · no GPU · 4 CPUs · 31 GB RAM · 4-bit by unsloth · 4.41 tok/s · $8.9491/MLinux · no GPU · 4 CPUs · 31 GB RAM · 3-bit by unsloth · 4.35 tok/s · $9.032/MLinux · no GPU · 4 CPUs · 31 GB RAM · 2-bit by unsloth · 4.17 tok/s · $9.4278/MLinux · no GPU · 4 CPUs · 31 GB RAM · 8-bit by unsloth · 4.04 tok/s · $9.856/MLinux · no GPU · 4 CPUs · 31 GB RAM · 8-bit by unsloth · 3.11 tok/s · $13.0639/M

Same build, different machines

Q4_K_M by unsloth on 6 machines: the fastest decodes 17× as fast as the slowest. One dot per machine, tokens per second.

  1. Linux, CPU only · 5 CPUs×173 tok/s
  2. 2× Tesla T4×261 tok/s
  3. NVIDIA L4×124 tok/s
  4. Linux, CPU only · 4 CPUs×24.4 tok/s

What it fits on

The smallest build needs about 12.9 GB; 9 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 cardnone
  3. NVIDIA GeForce GTX 16504 GB on the cardnone
  4. AMD Radeon(TM) Vega 8 Mobile Graphics7 GB RAMnone
  5. Windows, CPU only · 12 CPUs7 GB RAMnone
  6. Apple silicon8 GB RAMnone
  7. Apple M18 GB RAMnone
  8. Mac, CPU only · 8 CPUs8 GB RAMnone
  9. NVIDIA GeForce RTX 3080 Ti12 GB on the cardnone
  10. Tesla T415 GB on the card3 of 6
  11. Intel(R) UHD Graphics 73016 GB RAM3 of 6
  12. Intel(R) Iris(R) Xe Graphics16 GB RAM3 of 6
  13. Apple M416 GB RAM3 of 6
  14. Apple M116 GB RAM3 of 6
  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 128K. 2 walks ended because the model forgot, 13 because the machine could not hold the prompt.

machine · build4K8K16K32K64K128Kwhy it stopped
Gram ·a5be462× Tesla T4 · Q8_K_XL by unsloth128K recalled at every size asked
Gram ·dae679NVIDIA L4 · Q4_K_M by unsloth128K recalled at every size asked
Gram ·c76d302× Tesla T4 · Q4_K_M by unsloth64K the model lost the needle at 128K
Gram ·a5be462× Tesla T4 · Q3_K_S by unsloth64K the model lost the needle at 128K
Gram ·4e94b2Linux, CPU only · 4 CPUs · Q4_0 by unsloth4K too slow to answer here
Gram ·e61d3b2× Tesla T4 · Q4_K_M by unsloth4K too slow to answer here
Gram ·662a4eLinux, CPU only · 4 CPUs · Q3_K_S by unsloth4K too slow to answer here
Gram ·e3654fLinux, CPU only · 4 CPUs · Q4_K_M by unsloth4K too slow to answer here
Gram ·662a4eLinux, CPU only · 4 CPUs · Q4_0 by unsloth4K too slow to answer here
Gram ·e3654fLinux, CPU only · 4 CPUs · Q3_K_S by unsloth4K too slow to answer here
Gram ·662a4eLinux, CPU only · 4 CPUs · Q2_K by unsloth4K too slow to answer here
Gram ·662a4eLinux, CPU only · 4 CPUs · Q4_K_M by unsloth4K too slow to answer here
Gram ·e3654fLinux, CPU only · 4 CPUs · Q4_0 by unsloth4K too slow to answer here
Gram ·e3654fLinux, CPU only · 4 CPUs · Q2_K by unsloth4K too slow to answer here

The 14 longest of 17 walks; every one is on its build's page.

Carbon per million tokens

Lowest build on each machine · 3 of 7 read off a power sensor, the rest modelled from the parts · fewer grams is better

  1. Gram ·dae679477 g3.63 J/tok · sensor
  2. Gram ·a5be46556 g5.21 J/tok · sensor
  3. Gram ·e3654f589 g9.96 J/tok · modelled
  4. Gram ·e61d3b658 g11.12 J/tok · modelled
  5. Gram ·4e94b21.3 kg11.74 J/tok · modelled
  6. Gram ·c76d301.3 kg9.64 J/tok · sensor
  7. Gram ·662a4e1.7 kg12.91 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) · United States grid average · EU grid average (applied to NL) · world grid average (TW not in the table) · world grid average (Australia 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 · 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/gpt-oss-20b-GGUF:Q4_K_Mollamaquantizer · unsloth16.3 GB128K73.26
Q8_K_XL by unslothhf.co/unsloth/gpt-oss-20b-GGUF:Q8_K_XLollamaquantizer · unsloth19.6 GB128K44.22
Q3_K_S by unslothhf.co/unsloth/gpt-oss-20b-GGUF:Q3_K_Sollamaquantizer · unsloth16.2 GB64K54.73
Q4_0 by unslothhf.co/unsloth/gpt-oss-20b-GGUF:Q4_0ollamaquantizer · unsloth13.0 GB4K4.63
Q2_K by unslothhf.co/unsloth/gpt-oss-20b-GGUF:Q2_Kollamaquantizer · unsloth12.9 GB4K4.32
Q8_0 by unslothhf.co/unsloth/gpt-oss-20b-GGUF:Q8_0ollamaquantizer · unsloth13.7 GB4K4.02
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.

OpenAI-gpt-oss-20B-Claude-4.5-Opus-Heretic-Uncensored3 builds measuredgpt-oss-20b-Derestricted1 build measuredopenai_gpt-oss-20b-GGUF-MXFP4-Experimental1 build measuredHuihui-gpt-oss-20b-BF16-abliterated1 build measuredgpt-oss-20b-heretic1 build measuredOpenAI-gpt-oss-20B-INSTRUCT-Heretic-Uncensored-MXFP4not measuredgpt-oss-20b-MLX-6bit-gs32not measuredHuihui-gpt-oss-20b-BF16-abliterated-v2not measuredgpt-oss-20b-tq3not measuredgpt-oss-20b-OptiQ-4bitnot measuredInternVL3_5-GPT-OSS-20B-A4B-Preview-HFnot measuredgpt-oss-20b-unsloth-bnb-4bitnot measured