gpt-oss 20B

openai/gpt-oss-20b

3.6B active / 21B totalMoEText generationText input

OpenAI · 2 recorded configurations

Measured setup

Hardware

  • 2× RTX 3090 · GPU-resident

Runtime and tool

  • live llama-server

Quantisation

  • UD-Q8_K_XL

MTP draft depth

  • MTP off

Recorded setup

Every option opens a recorded setup. Where the corpus recorded nothing at this page's other settings, the option names the fields it switches with it. Allocation and runtime detail stay in the technical record below.

2× RTX 3090 · llama.cpp

UD-Q8_K_XL · MTP off

PP · left axisTG · right axis
View exact measurements
Exact context depth and measured fresh prefill, bulk prefix, and decode rates for unsloth/gpt-oss-20b-GGUF
Context depth (tokens)Fresh prefill PP (tok/s)TG (tok/s)
05,772.23168.95
2,0486,032.38165.54
4,0965,689.12161.35
8,1925,439.75152.13
16,3844,782.32140.69
32,7683,704.69117.75
65,5362,550.7788.86

Concurrent-stream measurements

Concurrent-stream throughput at the recorded existing-prefix depth for unsloth/gpt-oss-20b-GGUF
StreamsContext depth (tokens)Aggregate PP (tok/s)Aggregate TG (tok/s)TG / stream (tok/s)Aggregate all (tok/s)
105,772.23168.95168.95469.48
205,949.37267.76133.88730.47
305,761.65330.55110.18915.83
405,887.16340.9485.24961.44
505,615.81373.6674.731,061.16
605,521.12374.4062.401,072.24
705,506.78385.8355.121,111.05
805,514.35500.9862.621,433.80
12,0486,032.38165.54165.54461.95
22,0486,075.22260.65130.33714.47
32,0485,779.15312.06104.02870.93
42,0485,870.71322.8080.70913.84
52,0485,619.46351.5970.321,002.61
62,0485,625.07355.8359.301,023.40
72,0485,489.90366.2652.321,057.61
82,0485,469.01464.8658.111,336.94
14,0965,689.12161.35161.35449.30
24,0965,733.06247.13123.57677.31
34,0965,484.39295.9798.66826.05
44,0965,561.25304.4576.11862.44
54,0965,305.48328.4765.69937.73
64,0965,204.17334.7955.80962.30
74,0965,109.39342.5448.93989.22
84,0965,156.03432.8154.101,244.70
18,1925,439.75152.13152.13423.90
28,1925,559.53229.47114.73631.38
38,1925,218.43268.7389.58753.18
48,1925,395.34276.8869.22787.45
58,1925,018.45294.2358.85842.57
68,1924,993.92300.3450.06865.77
78,1924,904.03307.5143.93890.26
88,1924,909.78381.0547.631,098.65
116,3844,782.32140.69140.69391.09
216,3844,706.40201.99100.99554.29
316,3844,464.06228.2776.09639.54
416,3844,522.00231.5057.88658.92
516,3844,188.60245.6049.12703.42
616,3844,186.54247.4241.24713.40
716,3844,105.75250.5835.80725.36
816,3844,082.52301.2037.65869.13
132,7683,704.69117.75117.75326.06
232,7683,636.89155.8877.94428.45
332,7683,365.62172.1957.40482.98
432,7683,332.36171.1042.78487.19
532,7683,197.25179.7635.95516.02
632,7683,181.11177.6529.61513.42
732,7683,116.23177.6925.38515.71
832,7683,058.32201.0425.13582.86
165,5362,550.7788.8688.86244.89
265,5362,455.70106.9553.48293.99
365,5362,326.45114.2738.09321.68
465,5362,203.3795.9023.98275.69
565,5362,121.50112.5622.51324.10
665,5362,075.43109.4418.24317.12
765,5362,011.25107.7715.40313.64
865,5361,961.07112.4414.06327.65
Technical record and provenance

Run details

Source run ID
rtx3090x2-gpt-oss-20b-ud-q8-k-xl-2026-09-15-off
Measured date
2026-09-15
Hardware
2× NVIDIA GeForce RTX 3090
Runtime
llamaCpp · version: 0.3.0-dev (build 50, commit fc6545d) built with GNU 15.2.0 for Linux x86_64
Weight placement
gpu-resident
KV-cache precision
Q8_0
Flash Attention
Enabled
Batch size
512
Microbatch size
512
Prompt tokens
512
Generated tokens
256
Speculative decoding
None (standard decoding)
Measured repetitions
3
Notes
Automated llama-server KV-depth sweep (off). Existing depth is contextTokens; PP=512, TG=256, warmups=1, repetitions=3, Q8_0 K/V. Arithmetic mean of measured repetitions. PP/TG server durations are aligned to client completion timestamps for concurrent phase spans; aggregate-all includes SSH/HTTP wall time. Synthetic repeated-token prompts. Engine was reused as installed; no build was performed.
Source article
/benchmarks/
Raw log path
research/benchmarks/logs/rtx3090x2/2026-09-15-gpt-oss-20b-ud-q8-k-xl

Compare measured entries