Qwen3-8B · BF16

1× H100-80GB · 2026-09-24

Configuration

Model
Qwen/Qwen3-8B @ b968826d9c46
Precision
BF16
Max model length
16,384
Thinking mode
Off
Machine
gcp · a3-highgpu-1g (spot)
Engine
vLLM 0.30.0
Driver / CUDA
580.173.02 / 13.0
KV cache
377,024 tokens
Parallelism
TP1 PP1 DP1 EP1
Price
$6.57/h
Git commit
uncommitted-
Run date
2026-09-24
vllm serve Qwen/Qwen3-8B --revision main --max-model-len 16384 --gpu-memory-utilization 0.9 --port 8000 --seed 42 --dtype bfloat16 --tensor-parallel-size 1 --pipeline-parallel-size 1 --reasoning-parser qwen3 --default-chat-template-kwargs {"enable_thinking": false}

balanced-2048-2048

Max concurrent users: 40 within SLO (p99 TTFT ≤ 2.00 s, median inter-token ≤ 50 ms) · memory limit 92

UsersTTFT p95TTFT p99Inter-token medianInter-token p95Inter-token p99TPOT p99E2E p95E2E p99Output tok/stok/s/userTokens/J$/1M tokSLO
161.3 ms62.2 ms7 ms7.1 ms7.2 ms7 ms14.4 s14.4 s142142.90.33$12.82✓ pass
4203.7 ms205 ms7.5 ms7.7 ms7.8 ms7.6 ms15.6 s15.6 s506132.91.1$3.60✓ pass
16787.3 ms788.6 ms9.5 ms10.2 ms10.5 ms9.8 ms20.2 s20.3 s1,625105.73.22$1.12✓ pass
321.39 s1.59 s11.9 ms13.7 ms14.2 ms12.6 ms26.6 s26.8 s2,52984.34.41$0.722✓ pass
401.60 s1.99 s13.2 ms15.3 ms15.8 ms14.2 ms30.0 s30.4 s2,81575.54.79$0.648✓ pass
441.80 s2.21 s13.9 ms16 ms16.5 ms14.9 ms31.6 s32.0 s2,95071.94.89$0.619✕ fail
481.91 s2.36 s14.6 ms16.8 ms17.4 ms15.6 ms33.3 s33.7 s3,06768.65.08$0.595✕ fail
642.48 s3.09 s17.1 ms19.9 ms20.5 ms18.4 ms39.6 s40.2 s3,46358.65.28$0.527✕ fail
12812.4 s28.6 s25.5 ms26.9 ms202.8 ms43.4 ms85.8 s91.0 s3,91739.25.69$0.466✕ fail
25671.7 s106.4 s26.3 ms27.6 ms207.9 ms58.2 ms152.8 s169.1 s3,881385.64$0.470✕ fail
512201.8 s207.1 s26.3 ms27.7 ms207.7 ms58.1 ms296.7 s302.7 s3,901385.66$0.468✕ fail

chat-128-128

Max concurrent users: ≥ 512 within SLO (p99 TTFT ≤ 2.00 s, median inter-token ≤ 50 ms) · memory limit 1,472

UsersTTFT p95TTFT p99Inter-token medianInter-token p95Inter-token p99TPOT p99E2E p95E2E p99Output tok/stok/s/userTokens/J$/1M tokSLO
114.2 ms14.5 ms6.8 ms6.9 ms7 ms6.8 ms882.5 ms882.9 ms145146.50.34$12.56✓ pass
424.2 ms24.7 ms6.9 ms7 ms7.3 ms6.9 ms901.2 ms903.8 ms548144.51.3$3.33✓ pass
1657.5 ms64.7 ms7.1 ms7.3 ms8.1 ms7.4 ms966.7 ms971.5 ms2,156140.14.94$0.846✓ pass
3280.7 ms84.3 ms7.4 ms7.7 ms12.4 ms7.8 ms1.04 s1.05 s4,030135.28.2$0.453✓ pass
64181.9 ms189 ms8.1 ms8.7 ms14.6 ms9.2 ms1.24 s1.24 s6,958122.712.68$0.262✓ pass
128329 ms358.9 ms9.3 ms11.1 ms20.6 ms11.6 ms1.60 s1.61 s11,194107.117.77$0.163✓ pass
256828.5 ms916.9 ms12.8 ms16.3 ms144.8 ms17.9 ms2.59 s2.60 s13,09377.919.53$0.139✓ pass
5121.63 s1.89 s21.6 ms44.8 ms194.2 ms32.4 ms4.71 s4.74 s14,54046.321.42$0.126✓ pass

decode-128-1024

Max concurrent users: 396 within SLO (p99 TTFT ≤ 2.00 s, median inter-token ≤ 50 ms) · memory limit 327

UsersTTFT p95TTFT p99Inter-token medianInter-token p95Inter-token p99TPOT p99E2E p95E2E p99Output tok/stok/s/userTokens/J$/1M tokSLO
114.1 ms14.6 ms6.9 ms6.9 ms7 ms6.9 ms7.04 s7.05 s146145.80.34$12.53✓ pass
424.8 ms25.1 ms7 ms7.2 ms7.2 ms7 ms7.21 s7.21 s548142.71.29$3.33✓ pass
1649.2 ms49.5 ms7.5 ms7.9 ms8 ms7.5 ms7.76 s7.76 s2,117132.84.55$0.862✓ pass
3286 ms88.6 ms8.1 ms8.9 ms9.1 ms8.2 ms8.41 s8.41 s3,910123.37.43$0.467✓ pass
64249.2 ms251.1 ms9.5 ms11 ms11.4 ms9.6 ms9.99 s9.99 s6,597105.310.74$0.277✓ pass
128447 ms463.9 ms12 ms14.9 ms15.9 ms12.3 ms12.8 s12.8 s10,30183.314.94$0.177✓ pass
256848.8 ms929.8 ms18.1 ms23.6 ms25.3 ms18.7 ms19.5 s19.5 s13,56755.319.67$0.135✓ pass
3841.74 s1.75 s24.8 ms31.1 ms32.5 ms30 ms30.0 s31.4 s13,59140.419.96$0.134✓ pass
3921.76 s1.76 s25.7 ms31.7 ms49.1 ms31.7 ms31.7 s32.9 s13,3193919.58$0.137✓ pass
3961.95 s1.97 s25.8 ms31.8 ms48.2 ms32 ms32.1 s33.5 s13,32238.819.55$0.137✓ pass
4001.96 s2.13 s26 ms31.8 ms49.9 ms32.6 ms32.6 s33.9 s13,32338.419.57$0.137✕ fail
4162.33 s2.36 s27.5 ms32.9 ms52.8 ms35.2 ms35.3 s36.8 s13,06236.419.16$0.140✕ fail
4482.93 s5.43 s28.8 ms33.9 ms58.9 ms38.7 ms38.6 s40.6 s13,23034.719.31$0.138✕ fail
5124.33 s12.5 s30.4 ms38 ms71.1 ms46.2 ms46.2 s47.9 s13,52332.919.67$0.135✕ fail

longctx-8192-512

Max concurrent users: 7 within SLO (p99 TTFT ≤ 2.00 s, median inter-token ≤ 50 ms) · memory limit 43

UsersTTFT p95TTFT p99Inter-token medianInter-token p95Inter-token p99TPOT p99E2E p95E2E p99Output tok/stok/s/userTokens/J$/1M tokSLO
1253.9 ms254.6 ms7.3 ms7.4 ms7.5 ms7.3 ms3.98 s3.98 s1291370.29$14.16✓ pass
4893.4 ms898.1 ms8.6 ms8.7 ms8.9 ms9.9 ms5.30 s5.47 s378116.10.76$4.83✓ pass
71.35 s1.57 s10.2 ms10.5 ms10.9 ms12.6 ms6.96 s7.13 s50298.20.95$3.64✓ pass
101.91 s2.24 s11.8 ms12.4 ms25.3 ms15.5 ms8.79 s9.23 s62684.91.12$2.92✕ fail
162.92 s3.59 s13.8 ms14.7 ms223.5 ms19.8 ms12.2 s12.9 s78472.71.3$2.33✕ fail
325.62 s7.09 s20.3 ms39.6 ms231.2 ms33 ms21.6 s23.1 s95149.31.46$1.92✕ fail
6423.5 s26.5 s25.3 ms233 ms238.7 ms57.7 ms45.1 s48.2 s98239.61.48$1.86✕ fail
12847.0 s53.0 s25.3 ms233 ms238.5 ms57.9 ms68.6 s74.6 s99439.51.49$1.84✕ fail
256112.0 s118.3 s25.3 ms233.5 ms238.6 ms59.4 ms133.7 s139.9 s99439.51.49$1.84✕ fail
512245.5 s246.9 s25.3 ms233.4 ms238.8 ms59.4 ms267.1 s270.1 s99739.51.49$1.83✕ fail

prefill-1024-128

Max concurrent users: 80 within SLO (p99 TTFT ≤ 2.00 s, median inter-token ≤ 50 ms) · memory limit 327

UsersTTFT p95TTFT p99Inter-token medianInter-token p95Inter-token p99TPOT p99E2E p95E2E p99Output tok/stok/s/userTokens/J$/1M tokSLO
134.3 ms35.1 ms6.9 ms7 ms7.1 ms6.9 ms909.2 ms909.8 ms141145.20.32$12.94✓ pass
432.2 ms33 ms7.1 ms7.2 ms7.5 ms7.1 ms933.9 ms934.6 ms532140.51.24$3.43✓ pass
16236.1 ms261 ms7.9 ms8.1 ms19.7 ms10.2 ms1.36 s1.36 s1,757126.63.55$1.04✓ pass
32265.8 ms268.1 ms8.8 ms9.5 ms74.2 ms13.9 ms1.84 s1.84 s2,697113.14.77$0.677✓ pass
641.27 s1.55 s10.9 ms24.4 ms193.7 ms22 ms3.37 s3.76 s2,84791.64.31$0.641✓ pass
801.59 s1.95 s12 ms97.6 ms198.1 ms26 ms4.25 s4.63 s3,01883.64.53$0.605✓ pass
841.65 s2.04 s12.5 ms123.6 ms197.5 ms27.1 ms4.41 s4.79 s3,029804.56$0.603✕ fail
881.78 s2.17 s12.6 ms172.6 ms197.9 ms28 ms4.71 s5.11 s3,08779.54.6$0.591✕ fail
961.84 s2.33 s13.1 ms191.6 ms199.1 ms29.9 ms5.01 s5.56 s3,14176.54.62$0.581✕ fail
1282.50 s3.08 s14.9 ms195.4 ms200 ms37.6 ms6.51 s7.10 s3,34967.24.89$0.545✕ fail
2565.11 s6.33 s23.7 ms205.3 ms209 ms68.7 ms13.3 s14.5 s3,61842.25.24$0.504✕ fail
51213.5 s16.2 s31.4 ms211.5 ms214.8 ms113 ms25.0 s27.7 s3,51831.95.07$0.519✕ fail

Accuracy

TaskMetricScore± stderr
MMLU-Proexact_match63.2%0.4%
GPQA Diamondexact_match49.5%3.6%
GSM8Kexact_match83.9%1.0%
MATH-500math_verify80.8%1.8%
IFEvalprompt_level_strict_acc83.4%1.6%
ARC Challengeexact_match_letter88.0%1.0%