Qwen3-8B · FP8

1× H100-80GB · 2026-09-24

Configuration

Model
Qwen/Qwen3-8B @ b968826d9c46
Precision
FP8
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
430,272 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 --quantization fp8 --tensor-parallel-size 1 --pipeline-parallel-size 1 --reasoning-parser qwen3 --default-chat-template-kwargs {"enable_thinking": false}

balanced-2048-2048

Max concurrent users: 52 within SLO (p99 TTFT ≤ 2.00 s, median inter-token ≤ 50 ms) · memory limit 105

UsersTTFT p95TTFT p99Inter-token medianInter-token p95Inter-token p99TPOT p99E2E p95E2E p99Output tok/stok/s/userTokens/J$/1M tokSLO
150.3 ms50.7 ms4.8 ms4.9 ms5 ms4.8 ms9.88 s9.90 s208208.50.52$8.79✓ pass
4161.4 ms162.5 ms5.3 ms5.5 ms5.6 ms5.4 ms11.1 s11.1 s7161881.65$2.55✓ pass
16554.6 ms584.9 ms7.2 ms8 ms8.1 ms7.5 ms15.4 s15.5 s2,136138.64.32$0.854✓ pass
32996.6 ms1.17 s9.8 ms11.3 ms11.7 ms10.3 ms21.7 s21.9 s3,086101.95.39$0.591✓ pass
481.43 s1.77 s12.3 ms14.6 ms15.3 ms13 ms27.6 s27.9 s3,66781.65.84$0.498✓ pass
521.52 s1.93 s13.3 ms16 ms16.5 ms14.2 ms30.1 s30.5 s3,653755.83$0.500✓ pass
561.66 s2.08 s13.5 ms16.2 ms16.9 ms14.5 ms30.8 s31.3 s3,85573.95.93$0.473✕ fail
641.84 s2.36 s14.8 ms17.7 ms18.4 ms15.8 ms33.8 s34.3 s4,03367.56.09$0.453✕ fail
1284.23 s10.5 s24.6 ms27.3 ms153 ms34.8 ms68.2 s73.0 s4,42840.66.74$0.412✕ fail
25667.2 s71.1 s28.3 ms30 ms159 ms60.2 ms131.7 s135.6 s4,31035.46.5$0.423✕ fail
512196.0 s199.9 s28.3 ms29.9 ms159 ms60.2 ms259.9 s264.0 s4,34435.46.53$0.420✕ fail

chat-128-128

Max concurrent users: ≥ 512 within SLO (p99 TTFT ≤ 2.00 s, median inter-token ≤ 50 ms) · memory limit 1,680

UsersTTFT p95TTFT p99Inter-token medianInter-token p95Inter-token p99TPOT p99E2E p95E2E p99Output tok/stok/s/userTokens/J$/1M tokSLO
114.5 ms15 ms4.6 ms4.7 ms4.8 ms4.6 ms601.1 ms602.1 ms213216.50.55$8.56✓ pass
424.6 ms25.7 ms4.7 ms4.8 ms5 ms4.7 ms622.6 ms624.9 ms797212.12.09$2.29✓ pass
1656.1 ms57 ms4.9 ms5.2 ms6.1 ms5.1 ms676.1 ms680.6 ms3,088205.67.78$0.591✓ pass
3285.1 ms91.7 ms5.3 ms5.7 ms10.1 ms5.8 ms779 ms785.1 ms5,413188.511.33$0.337✓ pass
64155.3 ms161.6 ms5.9 ms6.8 ms14.6 ms6.9 ms946.3 ms953.5 ms9,17016918.33$0.199✓ pass
128330.6 ms333.9 ms7.5 ms9.4 ms19.1 ms9.3 ms1.31 s1.33 s13,377132.523.99$0.136✓ pass
256731.8 ms738 ms10.9 ms15.7 ms93.1 ms14.7 ms2.17 s2.18 s15,7049224.7$0.116✓ pass
5121.34 s1.51 s19.7 ms41.1 ms143.3 ms27.8 ms4.14 s4.16 s16,54250.825.55$0.110✓ pass

decode-128-1024

Max concurrent users: 472 within SLO (p99 TTFT ≤ 2.00 s, median inter-token ≤ 50 ms) · memory limit 373

UsersTTFT p95TTFT p99Inter-token medianInter-token p95Inter-token p99TPOT p99E2E p95E2E p99Output tok/stok/s/userTokens/J$/1M tokSLO
114.4 ms15 ms4.7 ms4.7 ms4.9 ms4.7 ms4.79 s4.80 s214214.50.54$8.52✓ pass
425.3 ms26 ms4.8 ms5 ms5.1 ms4.8 ms4.96 s4.96 s797207.71.99$2.29✓ pass
1657.6 ms57.7 ms5.3 ms5.6 ms5.8 ms5.3 ms5.43 s5.44 s3,027190.56.9$0.603✓ pass
3282 ms82.4 ms6 ms6.8 ms7 ms6.1 ms6.27 s6.27 s5,249165.810.15$0.348✓ pass
64201.9 ms212.8 ms7.3 ms8.7 ms9 ms7.4 ms7.65 s7.65 s8,621137.714.88$0.212✓ pass
128383.9 ms385.7 ms10.2 ms12.9 ms13.7 ms10.4 ms10.8 s10.8 s12,17698.419.5$0.150✓ pass
256694 ms738.3 ms16.1 ms21.4 ms23.4 ms16.6 ms17.3 s17.3 s15,2646222.47$0.120✓ pass
384964.2 ms1.10 s24.8 ms33 ms35.3 ms26.1 ms26.8 s26.9 s14,80340.221.57$0.123✓ pass
4481.56 s1.58 s27.1 ms34.4 ms59.5 ms33 ms32.9 s34.1 s14,78436.921.74$0.123✓ pass
4641.72 s1.73 s27.4 ms34.6 ms61.2 ms34.1 ms34.2 s35.4 s15,00636.522.11$0.122✓ pass
4721.87 s1.89 s27.4 ms34.7 ms63 ms34.8 ms34.7 s36.0 s15,10336.522.24$0.121✓ pass
4761.88 s2.02 s27.6 ms34.8 ms63.1 ms35.2 ms35.2 s36.3 s15,13136.222.29$0.121✕ fail
4802.02 s2.03 s27.8 ms34.8 ms64.4 ms35.6 ms35.6 s36.8 s15,1223622.27$0.121✕ fail
5122.48 s5.54 s28.3 ms36.2 ms69.4 ms38.2 ms38.4 s39.7 s15,38935.322.62$0.119✕ fail

longctx-8192-512

Max concurrent users: 10 within SLO (p99 TTFT ≤ 2.00 s, median inter-token ≤ 50 ms) · memory limit 49

UsersTTFT p95TTFT p99Inter-token medianInter-token p95Inter-token p99TPOT p99E2E p95E2E p99Output tok/stok/s/userTokens/J$/1M tokSLO
1205.1 ms206.1 ms5.1 ms5.2 ms5.4 ms5.1 ms2.81 s2.81 s183195.90.41$10.00✓ pass
4684.8 ms691.4 ms6.4 ms6.5 ms6.7 ms7.4 ms3.97 s3.98 s505155.61.01$3.61✓ pass
101.47 s1.71 s9.5 ms9.7 ms19.9 ms12.3 ms6.97 s7.30 s787105.21.44$2.32✓ pass
131.83 s2.21 s10.2 ms10.7 ms167.3 ms14 ms8.18 s8.62 s90397.71.56$2.02✕ fail
162.23 s2.74 s11.4 ms12.2 ms170.3 ms16.1 ms9.70 s10.2 s96687.51.6$1.89✕ fail
324.33 s5.46 s18.2 ms28.4 ms180.7 ms27.9 ms17.9 s19.1 s1,12454.81.75$1.62✕ fail
6416.3 s24.1 s27.2 ms184.9 ms187.4 ms56.5 ms37.8 s45.4 s1,14136.81.75$1.60✕ fail
12841.2 s48.2 s27.2 ms185.6 ms188.1 ms60.1 ms62.8 s69.7 s1,15436.71.76$1.58✕ fail
25698.0 s103.7 s27.2 ms185.6 ms188.2 ms62.2 ms122.5 s125.4 s1,15436.71.76$1.58✕ fail
512215.8 s215.9 s27.2 ms185.6 ms188.1 ms63.4 ms237.3 s237.5 s1,15736.71.76$1.58✕ fail

prefill-1024-128

Max concurrent users: 108 within SLO (p99 TTFT ≤ 2.00 s, median inter-token ≤ 50 ms) · memory limit 373

UsersTTFT p95TTFT p99Inter-token medianInter-token p95Inter-token p99TPOT p99E2E p95E2E p99Output tok/stok/s/userTokens/J$/1M tokSLO
129.3 ms31.1 ms4.7 ms4.8 ms4.9 ms4.7 ms625.8 ms626.9 ms205212.90.52$8.90✓ pass
431.8 ms32.2 ms4.9 ms5.1 ms5.3 ms4.9 ms654 ms655.5 ms760203.61.95$2.40✓ pass
16206.6 ms219.5 ms5.6 ms5.9 ms10.3 ms7.5 ms1.01 s1.01 s2,393178.15.22$0.763✓ pass
32260.1 ms287.2 ms6.7 ms9.3 ms60.8 ms10.7 ms1.44 s1.46 s3,444148.46.12$0.530✓ pass
64978.7 ms1.17 s8.6 ms19.5 ms142 ms16.9 ms2.57 s2.81 s3,701115.75.83$0.493✓ pass
961.40 s1.70 s11.1 ms135.9 ms144.3 ms23.3 ms3.85 s4.16 s4,02590.36.18$0.453✓ pass
1041.57 s1.85 s11.6 ms138.8 ms146.6 ms24.9 ms4.20 s4.48 s4,08086.26.29$0.447✓ pass
1081.59 s1.90 s11.8 ms139.4 ms146.5 ms25.5 ms4.30 s4.58 s4,129856.36$0.442✓ pass
1121.59 s2.01 s12 ms140.1 ms148.1 ms26.4 ms4.43 s4.85 s4,15383.36.38$0.439✕ fail
1281.85 s2.29 s12.9 ms142.7 ms148.6 ms29.3 ms5.09 s5.51 s4,26477.36.42$0.428✕ fail
2563.73 s4.68 s21.6 ms154.1 ms157 ms54.6 ms10.2 s11.1 s4,56146.26.67$0.400✕ fail
5127.72 s13.0 s35.1 ms162.8 ms167.1 ms105 ms19.0 s23.7 s4,41828.56.43$0.413✕ fail

Accuracy

TaskMetricScore± stderr
MMLU-Proexact_match62.6%0.4%
GPQA Diamondexact_match45.5%3.5%
GSM8Kexact_match86.4%0.9%
MATH-500math_verify82.2%1.7%
IFEvalprompt_level_strict_acc83.5%1.6%
ARC Challengeexact_match_letter87.6%1.0%