| model,format,tg128_ratio,tg128_ci95_low,tg128_ci95_high,pp512_ratio,pp512_ci95_low,pp512_ci95_high,batched_2seq_ratio,batched_4seq_ratio,batched_8seq_ratio,note,source |
| Qwen3-4B,IQ2_XXS (llama.cpp),1.000,,,1.000,,,1.00,1.00,1.00,reference; medians 197.1 tokens/s (tg128) and 7561 tokens/s (pp512),"paper: speed table (one RTX 5070 Ti Laptop GPU, 10 rotated rounds)" |
| Qwen3-4B,"Pingala-8 (B4, ours, shipped batch-1 kernel)",1.096,1.034,1.158,0.741,0.710,0.772,0.18,0.26,0.34,generation at batch 1 at least as fast as IQ2_XXS; prompt processing and batched decoding slower,"paper: speed table (one RTX 5070 Ti Laptop GPU, 10 rotated rounds)" |
| Qwen3-4B,"IQ2_KT, C4-calibrated (ik_llama.cpp)",0.995,0.952,1.038,0.949,0.906,0.992,0.66,0.73,0.73,ik_llama.cpp binary,"paper: speed table (one RTX 5070 Ti Laptop GPU, 10 rotated rounds)" |
| Qwen3-4B,Q4_K_M (llama.cpp),0.788,0.738,0.838,0.890,0.833,0.946,,,,larger 4-bit file,"paper: speed table (one RTX 5070 Ti Laptop GPU, 10 rotated rounds)" |
|
|
