Skip to content

Benchmark certificate · 2026-09-27.1

Both images passed their complete HTTP E2E suites before benchmarking. Prefill measures reading the prompt; decode measures generating the answer. Rates are tokens per second, with higher values indicating faster inference.

Certified prefill and decode results

Each chart row uses its own AVX512 result as the reference. Compare printed token rates across configurations, rather than bar lengths.

AVX512 certificate · AVX2 certificate · Combined results · SVG chart

Model / configuration AVX512 prefill AVX512 decode AVX2 prefill AVX2 decode
Qwen3.6-35B-A3B-UD-IQ3_S / 2×CPU · auto TP 294.7 37.0 111.7 28.0
Ornith-1.5-35B-Q4_K_M / 2×CPU · auto TP 352.4 42.0 124.3 32.2
Qwen3.5-122B-A10B-UD-Q8_K_XL-00001-of-00004 / 2×CPU + 2×CUDA · auto ExpertOverlay 124.9 14.0 79.7 13.9
Qwen3.5-122B-A10B-UD-Q8_K_XL-00001-of-00004 / 2×CPU + 2×CUDA + 4×ROCm · auto ExpertOverlay 124.6 20.1 124.6 19.7
Qwen3.5-122B-A10B-UD-Q8_K_XL-00001-of-00004 / 2×CPU + 2×ROCm · auto ExpertOverlay 164.7 22.6 102.1 20.9
Qwen3.5-122B-A10B-UD-Q8_K_XL-00001-of-00004 / 2×CPU + 4×ROCm · auto ExpertOverlay 170.8 27.5 147.0 26.8
Ornith-1.5-35B-Q4_K_M / 2×CUDA · auto TP 887.4 139.7 887.8 139.3
Qwen3.6-35B-A3B-UD-IQ3_S / 1×CUDA · auto single 1,737.6 283.0 1,761.0 287.1
Qwen3.8-27B-IQ4_XS / 2×CUDA · auto TP 468.1 50.4 469.4 50.3
Qwen3.8-27B-IQ4_XS / 2×CUDA · auto PP 743.7 40.0 760.7 40.3
Qwen3.8-27B-IQ4_XS / 1×CUDA · auto single 718.4 62.7 703.4 62.2
Qwen3.5-122B-A10B-UD-Q8_K_XL-00001-of-00004 / 2×CUDA + 4×ROCm · auto ExpertOverlay 153.6 25.0 155.2 24.8
Qwen3.8-27B-IQ4_XS / 2×CUDA + 2×ROCm · auto PP 301.7 32.0 303.1 32.2
Ornith-1.5-35B-Q4_K_M / 2×ROCm · auto TP 775.3 102.0 777.7 102.8
Qwen3.6-35B-A3B-UD-IQ3_S / 1×ROCm · auto single 830.9 148.5 835.4 148.6
Ornith-1.5-35B-Q4_K_M / 1×ROCm · auto single 821.8 166.2 817.8 166.0
Qwen3.8-27B-IQ4_XS / 2×ROCm · auto TP 247.1 39.1 247.5 39.2
Qwen3.8-27B-IQ4_XS / 2×ROCm · auto PP 184.8 22.4 185.7 21.6
Qwen3.8-27B-IQ4_XS / 1×ROCm · auto single 180.3 39.2 182.1 39.2