Correct HumanEval (harness bug) + publish Gemma 4 12B + quality leaderboard card

benchmarks.csv

@@ -13,111 +13,38 @@ Qwen3.6-27B,Dense,26.90,Q4_K_M,15.66,llama.cpp,CUDA,RTX 5090,32,pp4096,3644.93,2
 Qwen3.6-27B,Dense,26.90,Q4_K_M,15.66,llama.cpp,CUDA,RTX 5090,32,pp8192,3484.57,7.20,2026-05-28
 Qwen3.6-27B,Dense,26.90,Q4_K_M,15.66,llama.cpp,CUDA,RTX 5090,32,pp16384,3161.79,3.66,2026-05-28
 Qwen3.6-27B,Dense,26.90,Q4_K_M,15.66,llama.cpp,CUDA,RTX 5090,32,tg128,77.09,0.16,2026-05-28
-Nemotron-3-Nano-30B-A3B,MoE (3B active),31.58,Q4_K_M,22.88,llama.cpp,CUDA,RTX 5090,32,pp128,4423.05,74.78,2026-05-28
-Nemotron-3-Nano-30B-A3B,MoE (3B active),31.58,Q4_K_M,22.88,llama.cpp,CUDA,RTX 5090,32,pp512,10674.44,108.43,2026-05-28
-Nemotron-3-Nano-30B-A3B,MoE (3B active),31.58,Q4_K_M,22.88,llama.cpp,CUDA,RTX 5090,32,pp2048,10277.54,40.85,2026-05-28
-Nemotron-3-Nano-30B-A3B,MoE (3B active),31.58,Q4_K_M,22.88,llama.cpp,CUDA,RTX 5090,32,pp4096,9999.48,26.46,2026-05-28
-Nemotron-3-Nano-30B-A3B,MoE (3B active),31.58,Q4_K_M,22.88,llama.cpp,CUDA,RTX 5090,32,pp8192,9448.36,34.02,2026-05-28
-Nemotron-3-Nano-30B-A3B,MoE (3B active),31.58,Q4_K_M,22.88,llama.cpp,CUDA,RTX 5090,32,pp16384,8558.68,16.72,2026-05-28
-Nemotron-3-Nano-30B-A3B,MoE (3B active),31.58,Q4_K_M,22.88,llama.cpp,CUDA,RTX 5090,32,tg128,363.69,1.58,2026-05-28
-gpt-oss-20b,Dense,20.91,Q4_K_M,10.81,llama.cpp,CUDA,RTX 5090,32,pp128,7220.69,67.12,2026-05-28
-gpt-oss-20b,Dense,20.91,Q4_K_M,10.81,llama.cpp,CUDA,RTX 5090,32,pp512,16749.65,148.73,2026-05-28
-gpt-oss-20b,Dense,20.91,Q4_K_M,10.81,llama.cpp,CUDA,RTX 5090,32,pp2048,13524.44,12.42,2026-05-28
-gpt-oss-20b,Dense,20.91,Q4_K_M,10.81,llama.cpp,CUDA,RTX 5090,32,pp4096,11684.53,43.99,2026-05-28
-gpt-oss-20b,Dense,20.91,Q4_K_M,10.81,llama.cpp,CUDA,RTX 5090,32,pp8192,9413.7,16.38,2026-05-28
-gpt-oss-20b,Dense,20.91,Q4_K_M,10.81,llama.cpp,CUDA,RTX 5090,32,pp16384,6677.6,14.13,2026-05-28
-gpt-oss-20b,Dense,20.91,Q4_K_M,10.81,llama.cpp,CUDA,RTX 5090,32,tg128,367.9,1.18,2026-05-28
-Qwen3.6-27B-MTP,Dense (MTP),27.32,Q4_K_M,15.92,llama.cpp,CUDA,RTX 5090,32,pp128,2972.2,321.87,2026-05-28
-Qwen3.6-27B-MTP,Dense (MTP),27.32,Q4_K_M,15.92,llama.cpp,CUDA,RTX 5090,32,pp512,3835.77,43.26,2026-05-28
-Qwen3.6-27B-MTP,Dense (MTP),27.32,Q4_K_M,15.92,llama.cpp,CUDA,RTX 5090,32,pp2048,3746.68,1.53,2026-05-28
-Qwen3.6-27B-MTP,Dense (MTP),27.32,Q4_K_M,15.92,llama.cpp,CUDA,RTX 5090,32,pp4096,3655.53,9.44,2026-05-28
-Qwen3.6-27B-MTP,Dense (MTP),27.32,Q4_K_M,15.92,llama.cpp,CUDA,RTX 5090,32,pp8192,3495.59,4.04,2026-05-28
-Qwen3.6-27B-MTP,Dense (MTP),27.32,Q4_K_M,15.92,llama.cpp,CUDA,RTX 5090,32,pp16384,3161.77,3.81,2026-05-28
-Qwen3.6-27B-MTP,Dense (MTP),27.32,Q4_K_M,15.92,llama.cpp,CUDA,RTX 5090,32,tg128,76.99,0.09,2026-05-28
-Qwen3-Coder-Next,MoE,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp128,2381.32,29.12,2026-05-28
-Qwen3-Coder-Next,MoE,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp512,4447.3,39.42,2026-05-28
-Qwen3-Coder-Next,MoE,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp2048,4420.86,35.94,2026-05-28
-Qwen3-Coder-Next,MoE,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp4096,4380.75,11.49,2026-05-28
-Qwen3-Coder-Next,MoE,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp8192,4250.74,14.71,2026-05-28
-Qwen3-Coder-Next,MoE,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp16384,4042.73,18.93,2026-05-28
-Qwen3-Coder-Next,MoE,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,tg128,224.87,1.86,2026-05-28
-Qwen3.6-27B,Dense,26.90,Q6_K,20.97,llama.cpp,CUDA,RTX 5090,32,pp128,2560.42,230.77,2026-05-29
-Qwen3.6-27B,Dense,26.90,Q6_K,20.97,llama.cpp,CUDA,RTX 5090,32,pp512,3190.61,32.25,2026-05-29
-Qwen3.6-27B,Dense,26.90,Q6_K,20.97,llama.cpp,CUDA,RTX 5090,32,pp2048,3152.9,10.03,2026-05-29
-Qwen3.6-27B,Dense,26.90,Q6_K,20.97,llama.cpp,CUDA,RTX 5090,32,pp4096,3078.76,3.62,2026-05-29
-Qwen3.6-27B,Dense,26.90,Q6_K,20.97,llama.cpp,CUDA,RTX 5090,32,pp8192,2955.62,2.05,2026-05-29
-Qwen3.6-27B,Dense,26.90,Q6_K,20.97,llama.cpp,CUDA,RTX 5090,32,pp16384,2725.3,0.79,2026-05-29
-Qwen3.6-27B,Dense,26.90,Q6_K,20.97,llama.cpp,CUDA,RTX 5090,32,tg128,61.8,0.07,2026-05-29
-Qwen3.6-27B,Dense,26.90,Q6_K,20.97,llama.cpp,CUDA,RTX 5090,32,mmlu,87.92,,2026-05-29
-Qwen3.6-27B,Dense,26.90,Q6_K,20.97,llama.cpp,CUDA,RTX 5090,32,arc_challenge,96.93,,2026-05-29
-Qwen3.6-27B,Dense,26.90,Q6_K,20.97,llama.cpp,CUDA,RTX 5090,32,hellaswag,95.44,,2026-05-29
-Qwen3.6-27B,Dense,26.90,Q6_K,20.97,llama.cpp,CUDA,RTX 5090,32,gsm8k,97.27,,2026-05-29
-Qwen3.6-35B-A3B,MoE (3B active),34.66,UD-Q4_K_M,20.6,llama.cpp,CUDA,RTX 5090,32,pp128,3588.75,45.29,2026-05-29
-Qwen3.6-35B-A3B,MoE (3B active),34.66,UD-Q4_K_M,20.6,llama.cpp,CUDA,RTX 5090,32,pp512,9208.09,51.75,2026-05-29
-Qwen3.6-35B-A3B,MoE (3B active),34.66,UD-Q4_K_M,20.6,llama.cpp,CUDA,RTX 5090,32,pp2048,9014.19,41.42,2026-05-29
-Qwen3.6-35B-A3B,MoE (3B active),34.66,UD-Q4_K_M,20.6,llama.cpp,CUDA,RTX 5090,32,pp4096,8729.84,65.53,2026-05-29
-Qwen3.6-35B-A3B,MoE (3B active),34.66,UD-Q4_K_M,20.6,llama.cpp,CUDA,RTX 5090,32,pp8192,8362.15,25.66,2026-05-29
-Qwen3.6-35B-A3B,MoE (3B active),34.66,UD-Q4_K_M,20.6,llama.cpp,CUDA,RTX 5090,32,pp16384,7622.86,23.6,2026-05-29
-Qwen3.6-35B-A3B,MoE (3B active),34.66,UD-Q4_K_M,20.6,llama.cpp,CUDA,RTX 5090,32,tg128,270.9,1.47,2026-05-29
-Qwen3.6-35B-A3B,MoE (3B active),34.66,UD-Q4_K_M,20.6,llama.cpp,CUDA,RTX 5090,32,mmlu,84.99,,2026-05-29
-Qwen3.6-35B-A3B,MoE (3B active),34.66,UD-Q4_K_M,20.6,llama.cpp,CUDA,RTX 5090,32,arc_challenge,95.73,,2026-05-29
-Qwen3.6-35B-A3B,MoE (3B active),34.66,UD-Q4_K_M,20.6,llama.cpp,CUDA,RTX 5090,32,hellaswag,93.35,,2026-05-29
-Qwen3.6-35B-A3B,MoE (3B active),34.66,UD-Q4_K_M,20.6,llama.cpp,CUDA,RTX 5090,32,gsm8k,96.66,,2026-05-29
-Qwen3-Coder-Next,Dense,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp128,2377.56,29.85,2026-05-29
-Qwen3-Coder-Next,Dense,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp512,4433.02,39.13,2026-05-29
-Qwen3-Coder-Next,Dense,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp2048,4416.98,17.75,2026-05-29
-Qwen3-Coder-Next,Dense,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp4096,4372.29,6.73,2026-05-29
-Qwen3-Coder-Next,Dense,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp8192,4253.72,18.44,2026-05-29
-Qwen3-Coder-Next,Dense,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp16384,4022.08,7.21,2026-05-29
-Qwen3-Coder-Next,Dense,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,tg128,224.6,1.71,2026-05-29
-Qwen3-Coder-Next,Dense,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,mmlu,83.69,,2026-05-29
-Qwen3-Coder-Next,Dense,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,arc_challenge,95.99,,2026-05-29
-Qwen3-Coder-Next,Dense,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,hellaswag,89.32,,2026-05-29
-Qwen3-Coder-Next,Dense,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,gsm8k,95.98,,2026-05-29
-gemma-4-31B-it,Dense,30.70,Q6_K,23.46,llama.cpp,CUDA,RTX 5090,32,pp128,2485.54,170.26,2026-05-29
-gemma-4-31B-it,Dense,30.70,Q6_K,23.46,llama.cpp,CUDA,RTX 5090,32,pp512,2932.28,29.73,2026-05-29
-gemma-4-31B-it,Dense,30.70,Q6_K,23.46,llama.cpp,CUDA,RTX 5090,32,pp2048,2750.74,2.4,2026-05-29
-gemma-4-31B-it,Dense,30.70,Q6_K,23.46,llama.cpp,CUDA,RTX 5090,32,pp4096,2656.91,1.64,2026-05-29
-gemma-4-31B-it,Dense,30.70,Q6_K,23.46,llama.cpp,CUDA,RTX 5090,32,pp8192,2520.26,2.56,2026-05-29
-gemma-4-31B-it,Dense,30.70,Q6_K,23.46,llama.cpp,CUDA,RTX 5090,32,pp16384,2315.5,3.13,2026-05-29
-gemma-4-31B-it,Dense,30.70,Q6_K,23.46,llama.cpp,CUDA,RTX 5090,32,tg128,52.84,0.03,2026-05-29
-gemma-4-31B-it,Dense,30.70,Q6_K,23.46,llama.cpp,CUDA,RTX 5090,32,mmlu,87.82,,2026-05-29
-gemma-4-31B-it,Dense,30.70,Q6_K,23.46,llama.cpp,CUDA,RTX 5090,32,arc_challenge,97.61,,2026-05-29
-gemma-4-31B-it,Dense,30.70,Q6_K,23.46,llama.cpp,CUDA,RTX 5090,32,hellaswag,91.95,,2026-05-29
-gemma-4-31B-it,Dense,30.70,Q6_K,23.46,llama.cpp,CUDA,RTX 5090,32,gsm8k,97.5,,2026-05-29
-Nemotron-Cascade-2-30B-A3B,MoE (3B active),31.58,Q4_K_M,23.02,llama.cpp,CUDA,RTX 5090,32,pp128,4396.99,23.71,2026-05-29
-Nemotron-Cascade-2-30B-A3B,MoE (3B active),31.58,Q4_K_M,23.02,llama.cpp,CUDA,RTX 5090,32,pp512,10646.68,83.38,2026-05-29
-Nemotron-Cascade-2-30B-A3B,MoE (3B active),31.58,Q4_K_M,23.02,llama.cpp,CUDA,RTX 5090,32,pp2048,10338.38,32.81,2026-05-29
-Nemotron-Cascade-2-30B-A3B,MoE (3B active),31.58,Q4_K_M,23.02,llama.cpp,CUDA,RTX 5090,32,pp4096,10029.01,24.97,2026-05-29
-Nemotron-Cascade-2-30B-A3B,MoE (3B active),31.58,Q4_K_M,23.02,llama.cpp,CUDA,RTX 5090,32,pp8192,9508.0,20.63,2026-05-29
-Nemotron-Cascade-2-30B-A3B,MoE (3B active),31.58,Q4_K_M,23.02,llama.cpp,CUDA,RTX 5090,32,pp16384,8580.44,9.8,2026-05-29
-Nemotron-Cascade-2-30B-A3B,MoE (3B active),31.58,Q4_K_M,23.02,llama.cpp,CUDA,RTX 5090,32,tg128,350.84,1.4,2026-05-29
-Nemotron-Cascade-2-30B-A3B,MoE (3B active),31.58,Q4_K_M,23.02,llama.cpp,CUDA,RTX 5090,32,mmlu,74.42,,2026-05-29
-Nemotron-Cascade-2-30B-A3B,MoE (3B active),31.58,Q4_K_M,23.02,llama.cpp,CUDA,RTX 5090,32,arc_challenge,91.55,,2026-05-29
-Nemotron-Cascade-2-30B-A3B,MoE (3B active),31.58,Q4_K_M,23.02,llama.cpp,CUDA,RTX 5090,32,hellaswag,75.68,,2026-05-29
-Nemotron-Cascade-2-30B-A3B,MoE (3B active),31.58,Q4_K_M,23.02,llama.cpp,CUDA,RTX 5090,32,gsm8k,87.11,,2026-05-29
-Qwen3.6-27B,Dense,26.90,Q6_K,20.97,llama.cpp,CUDA,RTX 5090,32,pp32768,2332.97,9.07,2026-05-29
-Qwen3.6-27B,Dense,26.90,Q6_K,20.97,llama.cpp,CUDA,RTX 5090,32,pp65536,1772.23,0.55,2026-05-29
-Qwen3.6-35B-A3B,MoE (3B active),34.66,UD-Q4_K_M,20.6,llama.cpp,CUDA,RTX 5090,32,pp32768,6322.09,8.92,2026-05-29
-Qwen3.6-35B-A3B,MoE (3B active),34.66,UD-Q4_K_M,20.6,llama.cpp,CUDA,RTX 5090,32,pp65536,4685.37,3.87,2026-05-29
-Qwen3.6-35B-A3B,MoE (3B active),34.66,UD-Q4_K_M,20.6,llama.cpp,CUDA,RTX 5090,32,pp131072,2866.64,0.48,2026-05-29
-Qwen3-Coder-Next,Dense,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp32768,3561.75,2.06,2026-05-29
-Qwen3-Coder-Next,Dense,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp65536,2909.47,2.55,2026-05-29
-Nemotron-Cascade-2-30B-A3B,MoE (3B active),31.58,Q4_K_M,23.02,llama.cpp,CUDA,RTX 5090,32,pp32768,6968.23,12.54,2026-05-29
-Nemotron-Cascade-2-30B-A3B,MoE (3B active),31.58,Q4_K_M,23.02,llama.cpp,CUDA,RTX 5090,32,pp65536,4966.96,7.0,2026-05-29
-gpt-oss-120B,MoE (5.1B active),116.83,MXFP4,59.02,llama.cpp,CUDA,RTX 5090,32,pp128,124.5,33.1,2026-06-03
-gpt-oss-120B,MoE (5.1B active),116.83,MXFP4,59.02,llama.cpp,CUDA,RTX 5090,32,pp512,473.0,86.7,2026-06-03
-gpt-oss-120B,MoE (5.1B active),116.83,MXFP4,59.02,llama.cpp,CUDA,RTX 5090,32,pp2048,588.2,15.0,2026-06-03
-gpt-oss-120B,MoE (5.1B active),116.83,MXFP4,59.02,llama.cpp,CUDA,RTX 5090,32,tg128,46.68,0.16,2026-06-03
-gpt-oss-120B,MoE (5.1B active),116.83,MXFP4,59.02,llama.cpp,CUDA,RTX 5090,32,mmlu,89.47,,2026-06-03
-gpt-oss-120B,MoE (5.1B active),116.83,MXFP4,59.02,llama.cpp,CUDA,RTX 5090,32,arc_challenge,95.0,,2026-06-03
-gpt-oss-120B,MoE (5.1B active),116.83,MXFP4,59.02,llama.cpp,CUDA,RTX 5090,32,hellaswag,80.0,,2026-06-03
-gpt-oss-120B,MoE (5.1B active),116.83,MXFP4,59.02,llama.cpp,CUDA,RTX 5090,32,gsm8k,97.0,,2026-06-03
-gpt-oss-120B,MoE (5.1B active),116.83,MXFP4,59.02,llama.cpp,CUDA,RTX 5090,32,humaneval,98.0,,2026-06-03
-Qwen3.6-28B-REAP-A3B,MoE (3B active pruned),28.24,Q6_K,21.63,llama.cpp,CUDA,RTX 5090,32,pp512,8226.91,26.7,2026-06-03
-Qwen3.6-28B-REAP-A3B,MoE (3B active pruned),28.24,Q6_K,21.63,llama.cpp,CUDA,RTX 5090,32,tg128,246.62,1.7,2026-06-03
-Qwen3.6-28B-REAP-A3B,MoE (3B active pruned),28.24,Q6_K,21.63,llama.cpp,CUDA,RTX 5090,32,mmlu,87.72,,2026-06-03
-Qwen3.6-28B-REAP-A3B,MoE (3B active pruned),28.24,Q6_K,21.63,llama.cpp,CUDA,RTX 5090,32,arc_challenge,95.0,,2026-06-03
-Qwen3.6-28B-REAP-A3B,MoE (3B active pruned),28.24,Q6_K,21.63,llama.cpp,CUDA,RTX 5090,32,hellaswag,82.0,,2026-06-03
-Qwen3.6-28B-REAP-A3B,MoE (3B active pruned),28.24,Q6_K,21.63,llama.cpp,CUDA,RTX 5090,32,gsm8k,90.0,,2026-06-03
-Qwen3.6-28B-REAP-A3B,MoE (3B active pruned),28.24,Q6_K,21.63,llama.cpp,CUDA,RTX 5090,32,humaneval,94.0,,2026-06-03
+Nemotron-3-Nano-30B-A3B,MoE (3B active),31.58,Q4_K_M,22.88,llama.cpp,CUDA,RTX 5090,32,pp128,4423.05,74.78,2026-05-28
+Nemotron-3-Nano-30B-A3B,MoE (3B active),31.58,Q4_K_M,22.88,llama.cpp,CUDA,RTX 5090,32,pp512,10674.44,108.43,2026-05-28
+Nemotron-3-Nano-30B-A3B,MoE (3B active),31.58,Q4_K_M,22.88,llama.cpp,CUDA,RTX 5090,32,pp2048,10277.54,40.85,2026-05-28
+Nemotron-3-Nano-30B-A3B,MoE (3B active),31.58,Q4_K_M,22.88,llama.cpp,CUDA,RTX 5090,32,pp4096,9999.48,26.46,2026-05-28
+Nemotron-3-Nano-30B-A3B,MoE (3B active),31.58,Q4_K_M,22.88,llama.cpp,CUDA,RTX 5090,32,pp8192,9448.36,34.02,2026-05-28
+Nemotron-3-Nano-30B-A3B,MoE (3B active),31.58,Q4_K_M,22.88,llama.cpp,CUDA,RTX 5090,32,pp16384,8558.68,16.72,2026-05-28
+Nemotron-3-Nano-30B-A3B,MoE (3B active),31.58,Q4_K_M,22.88,llama.cpp,CUDA,RTX 5090,32,tg128,363.69,1.58,2026-05-28
+gpt-oss-20b,Dense,20.91,Q4_K_M,10.81,llama.cpp,CUDA,RTX 5090,32,pp128,7220.69,67.12,2026-05-28
+gpt-oss-20b,Dense,20.91,Q4_K_M,10.81,llama.cpp,CUDA,RTX 5090,32,pp512,16749.65,148.73,2026-05-28
+gpt-oss-20b,Dense,20.91,Q4_K_M,10.81,llama.cpp,CUDA,RTX 5090,32,pp2048,13524.44,12.42,2026-05-28
+gpt-oss-20b,Dense,20.91,Q4_K_M,10.81,llama.cpp,CUDA,RTX 5090,32,pp4096,11684.53,43.99,2026-05-28
+gpt-oss-20b,Dense,20.91,Q4_K_M,10.81,llama.cpp,CUDA,RTX 5090,32,pp8192,9413.7,16.38,2026-05-28
+gpt-oss-20b,Dense,20.91,Q4_K_M,10.81,llama.cpp,CUDA,RTX 5090,32,pp16384,6677.6,14.13,2026-05-28
+gpt-oss-20b,Dense,20.91,Q4_K_M,10.81,llama.cpp,CUDA,RTX 5090,32,tg128,367.9,1.18,2026-05-28
+Qwen3.6-27B-MTP,Dense (MTP),27.32,Q4_K_M,15.92,llama.cpp,CUDA,RTX 5090,32,pp128,2972.2,321.87,2026-05-28
+Qwen3.6-27B-MTP,Dense (MTP),27.32,Q4_K_M,15.92,llama.cpp,CUDA,RTX 5090,32,pp512,3835.77,43.26,2026-05-28
+Qwen3.6-27B-MTP,Dense (MTP),27.32,Q4_K_M,15.92,llama.cpp,CUDA,RTX 5090,32,pp2048,3746.68,1.53,2026-05-28
+Qwen3.6-27B-MTP,Dense (MTP),27.32,Q4_K_M,15.92,llama.cpp,CUDA,RTX 5090,32,pp4096,3655.53,9.44,2026-05-28
+Qwen3.6-27B-MTP,Dense (MTP),27.32,Q4_K_M,15.92,llama.cpp,CUDA,RTX 5090,32,pp8192,3495.59,4.04,2026-05-28
+Qwen3.6-27B-MTP,Dense (MTP),27.32,Q4_K_M,15.92,llama.cpp,CUDA,RTX 5090,32,pp16384,3161.77,3.81,2026-05-28
+Qwen3.6-27B-MTP,Dense (MTP),27.32,Q4_K_M,15.92,llama.cpp,CUDA,RTX 5090,32,tg128,76.99,0.09,2026-05-28
+Qwen3-Coder-Next,MoE,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp128,2381.32,29.12,2026-05-28
+Qwen3-Coder-Next,MoE,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp512,4447.3,39.42,2026-05-28
+Qwen3-Coder-Next,MoE,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp2048,4420.86,35.94,2026-05-28
+Qwen3-Coder-Next,MoE,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp4096,4380.75,11.49,2026-05-28
+Qwen3-Coder-Next,MoE,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp8192,4250.74,14.71,2026-05-28
+Qwen3-Coder-Next,MoE,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,pp16384,4042.73,18.93,2026-05-28
+Qwen3-Coder-Next,MoE,79.67,UD-Q2_K_XL,24.92,llama.cpp,CUDA,RTX 5090,32,tg128,224.87,1.86,2026-05-28
+Gemma 4 12B,Dense,11.91,Q6_K,9.1,llama.cpp,CUDA,RTX 5090,32,pp128,5099.28,452.85,2026-06-04
+Gemma 4 12B,Dense,11.91,Q6_K,9.1,llama.cpp,CUDA,RTX 5090,32,pp512,7160.37,149.07,2026-06-04
+Gemma 4 12B,Dense,11.91,Q6_K,9.1,llama.cpp,CUDA,RTX 5090,32,pp2048,6788.28,10.42,2026-06-04
+Gemma 4 12B,Dense,11.91,Q6_K,9.1,llama.cpp,CUDA,RTX 5090,32,pp4096,6605.39,1.81,2026-06-04
+Gemma 4 12B,Dense,11.91,Q6_K,9.1,llama.cpp,CUDA,RTX 5090,32,pp8192,6359.06,5.49,2026-06-04
+Gemma 4 12B,Dense,11.91,Q6_K,9.1,llama.cpp,CUDA,RTX 5090,32,pp16384,5846.08,5.21,2026-06-04
+Gemma 4 12B,Dense,11.91,Q6_K,9.1,llama.cpp,CUDA,RTX 5090,32,tg128,122.3,0.19,2026-06-04

reports/gemma-4-12b-it-q6-k.md

@@ -0,0 +1,97 @@
+# Benchmark Report: gemma-4-12b-it (Q6_K)
+
+**Date:** 2026-06-04  
+**Author:** WITCHEER  
+**Platform:** NVIDIA GeForce RTX 5090 Benchmark Rig (capsule)  
+
+---
+
+## Model
+
+| Field | Value |
+|-------|-------|
+| Model | gemma-4-12b-it |
+| Parameters | 11.91 B (dense) |
+| Quantization | Q6_K |
+| File size | 9.11 GiB |
+| Engine | llama.cpp (CUDA 12.8 (patched)) |
+
+## Hardware
+
+| Component | Spec |
+|-----------|------|
+| GPU | NVIDIA GeForce RTX 5090 |
+| CPU | AMD Ryzen 5 9600 |
+| RAM | 64GB DDR5-5600 |
+| OS | Ubuntu 26.04 LTS |
+| CUDA | 12.8 (patched) |
+
+---
+
+## Quality Benchmarks
+
+All benchmarks use generative evaluation via llama-server chat completions. Multiple-choice tasks (MMLU, ARC, HellaSwag) use letter extraction instead of loglikelihood scoring -- results are internally consistent for model comparison but absolute scores may differ from logprob-based evaluations by 5-15%.
+
+### Summary
+
+| Benchmark | Score | Metric |
+|-----------|------:|--------|
+| **MMLU** | **78.86%** | accuracy |
+| **ARC-Challenge** | **94.03%** | accuracy |
+| **HellaSwag** | **81.62%** | accuracy |
+| **HumanEval** | **87.20%** | pass@1 |
+| **GSM8K** | **96.36%** | exact_match |
+
+### MMLU Breakdown by Category
+
+| Category | Score | Correct / Total |
+|----------|------:|----------------:|
+| Stem | 77.84% | 1,173 / 1,507 |
+| Humanities | 77.81% | 1,231 / 1,582 |
+| Social Sciences | 88.57% | 1,464 / 1,653 |
+| Other | 73.19% | 1,660 / 2,268 |
+
+*Sampled at 50% (seed 42)*
+
+---
+
+## Speed Benchmarks
+
+Measured with `llama-bench`. All layers GPU-offloaded (`-ngl 99`).
+
+### Prompt Processing (tokens/s)
+
+| Context Length | Speed | +/-sigma |
+|---------------:|------:|---------:|
+| 128 | 5,099 | 452.9 |
+| 512 | 7,160 | 149.1 |
+| 2048 | 6,788 | 10.4 |
+| 4096 | 6,605 | 1.8 |
+| 8192 | 6,359 | 5.5 |
+| 16384 | 5,846 | 5.2 |
+
+### Generation (tokens/s)
+
+| Metric | Speed | +/-sigma |
+|--------|------:|---------:|
+| tg128 | 122.3 | 0.2 |
+
+---
+
+## Methodology
+
+### Evaluation Framework
+
+Custom generative evaluators built for this rig. All benchmarks run through llama-server's `/v1/chat/completions` endpoint.
+
+- **Scoring:** Generative evaluation (not loglikelihood)
+- **Thinking:** disabled
+- **MCQ scoring:** First valid letter extracted from response (A/B/C/D)
+- **Sampling:** 50% of dataset used
+- **Temperature:** 0 (deterministic)
+- **Max tokens:** 2,048
+- **GPU offload:** All layers (`-ngl 99`)
+
+---
+
+*Benchmarked by WITCHEER on the RTX 5090 Benchmark Rig. Source: [github.com/notwitcheer/llm-bench-rig/blob/main/reports/gemma-4-12b-it-q6-k.md](https://github.com/notwitcheer/llm-bench-rig/blob/main/reports/gemma-4-12b-it-q6-k.md). Dataset: [huggingface.co/datasets/witcheer/rtx-5090-benchmarks/blob/main/reports/gemma-4-12b-it-q6-k.md](https://huggingface.co/datasets/witcheer/rtx-5090-benchmarks/blob/main/reports/gemma-4-12b-it-q6-k.md).*

reports/gemma-4-31b-it-q6-k.md

@@ -1,6 +1,6 @@
 # Benchmark Report: gemma-4-31B-it (Q6_K)
 
-**Date:** 2026-05-29  
+**Date:** 2026-06-04  
 **Author:** WITCHEER  
 **Platform:** NVIDIA GeForce RTX 5090 Benchmark Rig (capsule)  
 
@@ -39,7 +39,7 @@ All benchmarks use generative evaluation via llama-server chat completions. Mult
 | **MMLU** | **87.82%** | accuracy |
 | **ARC-Challenge** | **97.61%** | accuracy |
 | **HellaSwag** | **91.95%** | accuracy |
-| **HumanEval** | **95.73%** | pass@1 |
+| **HumanEval** | **96.34%** | pass@1 |
 | **GSM8K** | **97.50%** | exact_match |
 
 ### MMLU Breakdown by Category
@@ -63,18 +63,18 @@ Measured with `llama-bench`. All layers GPU-offloaded (`-ngl 99`).
 
 | Context Length | Speed | +/-sigma |
 |---------------:|------:|---------:|
-| 128 | 2,486 | 170.3 |
-| 512 | 2,932 | 29.7 |
-| 2048 | 2,751 | 2.4 |
-| 4096 | 2,657 | 1.6 |
-| 8192 | 2,520 | 2.6 |
-| 16384 | 2,316 | 3.1 |
+| 128 | 2,509 | 127.0 |
+| 512 | 2,939 | 25.6 |
+| 2048 | 2,752 | 1.5 |
+| 4096 | 2,660 | 1.9 |
+| 8192 | 2,525 | 1.3 |
+| 16384 | 2,325 | 1.6 |
 
 ### Generation (tokens/s)
 
 | Metric | Speed | +/-sigma |
 |--------|------:|---------:|
-| tg128 | 52.8 | 0.0 |
+| tg128 | 53.0 | 0.0 |
 
 ---
 

reports/gpt-oss-20b-q4-k-m.md

@@ -1,8 +1,8 @@
 # Benchmark Report: gpt-oss-20b (Q4_K_M)
 
-**Date:** 2026-05-28  
+**Date:** 2026-06-04  
 **Author:** WITCHEER  
-**Platform:** RTX 5090 Benchmark Rig (capsule)  
+**Platform:** NVIDIA GeForce RTX 5090 Benchmark Rig (capsule)  
 
 ---
 
@@ -11,82 +11,46 @@
 | Field | Value |
 |-------|-------|
 | Model | gpt-oss-20b |
-| Parameters | 20.91B (dense) |
+| Parameters | 20.91 B (dense) |
 | Quantization | Q4_K_M |
 | File size | 10.83 GiB |
-| Engine | llama.cpp (CUDA 12.8, sm_120) |
+| Engine | llama.cpp (CUDA 12.8 (patched)) |
 
 ## Hardware
 
 | Component | Spec |
 |-----------|------|
-| GPU | NVIDIA GeForce RTX 5090 (32 GB GDDR7) |
-| CPU | AMD Ryzen 5 9600 (6c/12t, 3.8/5.2 GHz) |
-| RAM | 64 GB DDR5-5600 |
-| OS | Ubuntu Server 26.04 LTS (headless) |
-| CUDA | 12.8 (patched for glibc 2.41 compat) |
+| GPU | NVIDIA GeForce RTX 5090 |
+| CPU | AMD Ryzen 5 9600 |
+| RAM | 64GB DDR5-5600 |
+| OS | Ubuntu 26.04 LTS |
+| CUDA | 12.8 (patched) |
 
 ---
 
 ## Quality Benchmarks
 
-All benchmarks use generative evaluation via llama-server chat completions. Multiple-choice tasks (MMLU, ARC, HellaSwag) use letter extraction instead of loglikelihood scoring — results are internally consistent for model comparison but absolute scores may differ from logprob-based evaluations by 5–15%.
+All benchmarks use generative evaluation via llama-server chat completions. Multiple-choice tasks (MMLU, ARC, HellaSwag) use letter extraction instead of loglikelihood scoring -- results are internally consistent for model comparison but absolute scores may differ from logprob-based evaluations by 5-15%.
 
 ### Summary
 
-| Benchmark | Score | Metric | Correct / Total | Time |
-|-----------|------:|--------|----------------:|-----:|
-| **MMLU** (5-shot) | **78.56%** | accuracy | 11,031 / 14,042 | 3h 49m |
-| **ARC-Challenge** (25-shot) | **94.62%** | accuracy | 1,109 / 1,172 | 10m 40s |
-| **HellaSwag** (10-shot) | **74.49%** | accuracy | 7,480 / 10,042 | 3h 31m |
-| **GSM8K** (5-shot, CoT) | **94.77%** | exact match | 1,250 / 1,319 | 22m 0s |
-| **HumanEval** (0-shot) | **12.20%** | pass@1 | 20 / 164 | 2m 48s |
-
-**Total evaluation time:** 7h 56m
+| Benchmark | Score | Metric |
+|-----------|------:|--------|
+| **MMLU** | **78.56%** | accuracy |
+| **ARC-Challenge** | **94.62%** | accuracy |
+| **HellaSwag** | **74.49%** | accuracy |
+| **HumanEval** | **94.51%** | pass@1 |
+| **GSM8K** | **94.77%** | exact_match |
 
 ### MMLU Breakdown by Category
 
 | Category | Score | Correct / Total |
 |----------|------:|----------------:|
-| STEM | 89.83% | 2,711 / 3,018 |
-| Social Sciences | 84.45% | 2,796 / 3,311 |
+| Stem | 89.83% | 2,711 / 3,018 |
 | Humanities | 77.45% | 2,456 / 3,171 |
+| Social Sciences | 84.45% | 2,796 / 3,311 |
 | Other | 67.55% | 3,068 / 4,542 |
 
-**Top 5 subjects:**
-
-| Subject | Score |
-|---------|------:|
-| High School Computer Science | 99.0% |
-| Elementary Mathematics | 96.8% |
-| College Physics | 95.1% |
-| High School Mathematics | 93.7% |
-| College Biology | 92.4% |
-
-**Bottom 5 subjects:**
-
-| Subject | Score |
-|---------|------:|
-| Professional Law | 44.3% |
-| Global Facts | 47.0% |
-| Virology | 59.0% |
-| Moral Disputes | 67.9% |
-| Philosophy | 68.8% |
-
-### Parse Reliability
-
-The model uses extended reasoning (`reasoning_content` field) before responding. With `max_tokens=2048`, most reasoning chains complete successfully.
-
-| Benchmark | Parse Failures | Failure Rate |
-|-----------|---------------:|-------------:|
-| MMLU | 653 | 4.6% |
-| ARC-Challenge | 5 | 0.4% |
-| HellaSwag | 37 | 0.4% |
-| GSM8K | 0 | 0.0% |
-| **Total** | **695** | **2.6%** |
-
-Parse failures are scored as incorrect. The majority occur in MMLU subjects with long reasoning chains (professional_law, moral_scenarios) where the model's thinking exceeds the token budget.
-
 ---
 
 ## Speed Benchmarks
@@ -95,24 +59,20 @@ Measured with `llama-bench`. All layers GPU-offloaded (`-ngl 99`).
 
 ### Prompt Processing (tokens/s)
 
-| Context Length | Speed | ±σ |
-|---------------:|------:|---:|
-| 128 | 7,221 | 67 |
-| 512 | 16,750 | 149 |
-| 2,048 | 13,524 | 12 |
-| 4,096 | 11,685 | 44 |
-| 8,192 | 9,414 | 16 |
-| 16,384 | 6,678 | 14 |
+| Context Length | Speed | +/-sigma |
+|---------------:|------:|---------:|
+| 128 | 7,172 | 73.0 |
+| 512 | 16,646 | 117.8 |
+| 2048 | 13,456 | 36.4 |
+| 4096 | 11,684 | 28.1 |
+| 8192 | 9,408 | 22.7 |
+| 16384 | 6,669 | 4.6 |
 
 ### Generation (tokens/s)
 
-| Metric | Speed | ±σ |
-|--------|------:|---:|
-| tg128 | 367.9 | 1.2 |
-
-### Context Degradation
-
-Prompt processing peaks at 512 tokens (16,750 t/s) then drops 60% at 16K context (6,678 t/s). This is the steepest degradation of any model in the rig — characteristic of smaller dense models with limited KV-cache efficiency.
+| Metric | Speed | +/-sigma |
+|--------|------:|---------:|
+| tg128 | 367.4 | 0.9 |
 
 ---
 
@@ -120,50 +80,16 @@ Prompt processing peaks at 512 tokens (16,750 t/s) then drops 60% at 16K context
 
 ### Evaluation Framework
 
-Custom generative evaluators built for this rig. No dependency on `lm-evaluation-harness` — all benchmarks run through llama-server's `/v1/chat/completions` endpoint.
+Custom generative evaluators built for this rig. All benchmarks run through llama-server's `/v1/chat/completions` endpoint.
 
-| Benchmark | Dataset | Eval Split | Few-shot | Scoring |
-|-----------|---------|-----------|----------|---------|
-| MMLU | `cais/mmlu` | test (14,042) | 5-shot per subject from `dev` | First valid A/B/C/D letter extracted from response |
-| ARC-Challenge | `allenai/ai2_arc` | test (1,172) | 25-shot from `train` | First valid letter, numeric labels normalized to A–D |
-| HellaSwag | `Rowan/hellaswag` | validation (10,042) | 10-shot from `train` | First valid A/B/C/D letter |
-| GSM8K | `openai/gsm8k` | test (1,319) | 5-shot CoT from `train` | Exact match on extracted numeric answer |
-| HumanEval | `openai/openai_humaneval` | test (164) | 0-shot | pass@1 via subprocess execution (10s timeout) |
-
-### Inference Configuration
-
-- **Server:** llama-server (llama.cpp, CUDA 12.8, Blackwell sm_120)
+- **Scoring:** Generative evaluation (not loglikelihood)
+- **Thinking:** disabled
+- **MCQ scoring:** First valid letter extracted from response (A/B/C/D)
+- **Sampling:** 50% of dataset used
 - **Temperature:** 0 (deterministic)
-- **Max tokens:** 2,048 (accommodates reasoning models)
+- **Max tokens:** 2,048
 - **GPU offload:** All layers (`-ngl 99`)
-- **Serving:** Single request, sequential (no batching)
-
-### Differences from Standard Benchmarks
-
-- **Generative vs loglikelihood:** MMLU, ARC, and HellaSwag are traditionally scored using token logprobabilities. This rig uses generative letter extraction, which typically yields scores 5–15% lower on the same model. Rankings between models remain consistent.
-- **Thinking models:** gpt-oss-20b produces extended reasoning in a separate `reasoning_content` field. When the primary `content` field is empty, the evaluator falls back to parsing the reasoning chain for the final answer.
-- **No normalized accuracy:** Standard HellaSwag reporting uses `acc_norm` (length-normalized). This rig reports raw accuracy, which may be lower for completions of varying length.
-
----
-
-## Reproduction
-
-```bash
-# On capsule (192.168.1.9)
-cd ~/benchmark-rig && source venv/bin/activate
-
-# Full benchmark (speed + quality)
-python3 bench.py /path/to/gpt-oss-20b-Q4_K_M.gguf
-
-# Quality only
-python3 bench.py /path/to/gpt-oss-20b-Q4_K_M.gguf --quality-only
-
-# Individual evaluator
-python3 -m lib.evals.mmlu --api-base http://127.0.0.1:8090/v1 --model gpt-oss-20b
-```
-
-All results, detailed per-subject breakdowns, and checkpoint files are stored in `results/gpt-oss-20b-q4-k-m/`.
 
 ---
 
-*Benchmarked by WITCHEER on the RTX 5090 Benchmark Rig. Source: [github.com/notwitcheer/llm-bench-rig](https://github.com/notwitcheer/llm-bench-rig). Dataset: [huggingface.co/datasets/witcheer/rtx-5090-benchmarks](https://huggingface.co/datasets/witcheer/rtx-5090-benchmarks).*
+*Benchmarked by WITCHEER on the RTX 5090 Benchmark Rig. Source: [github.com/notwitcheer/llm-bench-rig/blob/main/reports/gpt-oss-20b-q4-k-m.md](https://github.com/notwitcheer/llm-bench-rig/blob/main/reports/gpt-oss-20b-q4-k-m.md). Dataset: [huggingface.co/datasets/witcheer/rtx-5090-benchmarks/blob/main/reports/gpt-oss-20b-q4-k-m.md](https://huggingface.co/datasets/witcheer/rtx-5090-benchmarks/blob/main/reports/gpt-oss-20b-q4-k-m.md).*

reports/qwen3-6-27b-q6-k.md

@@ -1,6 +1,6 @@
 # Benchmark Report: Qwen3.6-27B (Q6_K)
 
-**Date:** 2026-05-29  
+**Date:** 2026-06-04  
 **Author:** WITCHEER  
 **Platform:** NVIDIA GeForce RTX 5090 Benchmark Rig (capsule)  
 
@@ -39,7 +39,7 @@ All benchmarks use generative evaluation via llama-server chat completions. Mult
 | **MMLU** | **87.92%** | accuracy |
 | **ARC-Challenge** | **96.93%** | accuracy |
 | **HellaSwag** | **95.44%** | accuracy |
-| **HumanEval** | **18.90%** | pass@1 |
+| **HumanEval** | **92.68%** | pass@1 |
 | **GSM8K** | **97.27%** | exact_match |
 
 ### MMLU Breakdown by Category
@@ -63,20 +63,18 @@ Measured with `llama-bench`. All layers GPU-offloaded (`-ngl 99`).
 
 | Context Length | Speed | +/-sigma |
 |---------------:|------:|---------:|
-| 128 | 2,560 | 230.8 |
-| 512 | 3,191 | 32.2 |
-| 2048 | 3,153 | 10.0 |
-| 4096 | 3,079 | 3.6 |
-| 8192 | 2,956 | 2.0 |
-| 16384 | 2,725 | 0.8 |
-| 32768 | 2,333 | 9.1 |
-| 65536 | 1,772 | 0.6 |
+| 128 | 2,580 | 235.1 |
+| 512 | 3,222 | 38.0 |
+| 2048 | 3,187 | 1.9 |
+| 4096 | 3,115 | 3.4 |
+| 8192 | 2,992 | 1.5 |
+| 16384 | 2,751 | 2.3 |
 
 ### Generation (tokens/s)
 
 | Metric | Speed | +/-sigma |
 |--------|------:|---------:|
-| tg128 | 61.8 | 0.1 |
+| tg128 | 61.9 | 0.1 |
 
 ---
 

reports/qwen3-6-35b-a3b-ud-q4-k-m.md

@@ -1,6 +1,6 @@
 # Benchmark Report: Qwen3.6-35B-A3B (UD-Q4_K_M)
 
-**Date:** 2026-05-29  
+**Date:** 2026-06-04  
 **Author:** WITCHEER  
 **Platform:** NVIDIA GeForce RTX 5090 Benchmark Rig (capsule)  
 
@@ -39,7 +39,7 @@ All benchmarks use generative evaluation via llama-server chat completions. Mult
 | **MMLU** | **84.99%** | accuracy |
 | **ARC-Challenge** | **95.73%** | accuracy |
 | **HellaSwag** | **93.35%** | accuracy |
-| **HumanEval** | **37.20%** | pass@1 |
+| **HumanEval** | **95.73%** | pass@1 |
 | **GSM8K** | **96.66%** | exact_match |
 
 ### MMLU Breakdown by Category
@@ -63,21 +63,18 @@ Measured with `llama-bench`. All layers GPU-offloaded (`-ngl 99`).
 
 | Context Length | Speed | +/-sigma |
 |---------------:|------:|---------:|
-| 128 | 3,589 | 45.3 |
-| 512 | 9,208 | 51.8 |
-| 2048 | 9,014 | 41.4 |
-| 4096 | 8,730 | 65.5 |
-| 8192 | 8,362 | 25.7 |
-| 16384 | 7,623 | 23.6 |
-| 32768 | 6,322 | 8.9 |
-| 65536 | 4,685 | 3.9 |
-| 131072 | 2,867 | 0.5 |
+| 128 | 3,584 | 45.7 |
+| 512 | 9,217 | 50.7 |
+| 2048 | 9,004 | 47.0 |
+| 4096 | 8,731 | 56.7 |
+| 8192 | 8,347 | 21.4 |
+| 16384 | 7,641 | 10.8 |
 
 ### Generation (tokens/s)
 
 | Metric | Speed | +/-sigma |
 |--------|------:|---------:|
-| tg128 | 270.9 | 1.5 |
+| tg128 | 270.6 | 1.7 |
 
 ---
 

reports/qwen3-coder-next-ud-q2-k-xl.md

@@ -1,6 +1,6 @@
 # Benchmark Report: Qwen3-Coder-Next (UD-Q2_K_XL)
 
-**Date:** 2026-05-29  
+**Date:** 2026-06-04  
 **Author:** WITCHEER  
 **Platform:** NVIDIA GeForce RTX 5090 Benchmark Rig (capsule)  
 
@@ -39,7 +39,7 @@ All benchmarks use generative evaluation via llama-server chat completions. Mult
 | **MMLU** | **83.69%** | accuracy |
 | **ARC-Challenge** | **95.99%** | accuracy |
 | **HellaSwag** | **89.32%** | accuracy |
-| **HumanEval** | **10.37%** | pass@1 |
+| **HumanEval** | **93.29%** | pass@1 |
 | **GSM8K** | **95.98%** | exact_match |
 
 ### MMLU Breakdown by Category
@@ -63,20 +63,18 @@ Measured with `llama-bench`. All layers GPU-offloaded (`-ngl 99`).
 
 | Context Length | Speed | +/-sigma |
 |---------------:|------:|---------:|
-| 128 | 2,378 | 29.9 |
-| 512 | 4,433 | 39.1 |
-| 2048 | 4,417 | 17.8 |
-| 4096 | 4,372 | 6.7 |
-| 8192 | 4,254 | 18.4 |
-| 16384 | 4,022 | 7.2 |
-| 32768 | 3,562 | 2.1 |
-| 65536 | 2,909 | 2.5 |
+| 128 | 2,387 | 29.0 |
+| 512 | 4,439 | 37.6 |
+| 2048 | 4,432 | 21.0 |
+| 4096 | 4,369 | 10.4 |
+| 8192 | 4,275 | 21.3 |
+| 16384 | 4,024 | 9.7 |
 
 ### Generation (tokens/s)
 
 | Metric | Speed | +/-sigma |
 |--------|------:|---------:|
-| tg128 | 224.6 | 1.7 |
+| tg128 | 224.7 | 1.5 |
 
 ---