study,semafold,benchmark_version,0.1.0,library_version,0.2.0,reviewed,2026-09-25

Reconstruction-Oriented Storage (`TurboQuantMSE`)
Setting,Dense float32,Dense fp16/bf16,Semafold Artifact,Saved vs float32,Smaller vs float32,Smaller vs fp16,Ratio vs float32,Ratio vs fp16,Quality Proxy
2-bit MSE,"786,432 B","393,216 B","50,077 B","736,355 B",93.63%,87.26%,15.70x,7.85x,observed MSE 0.1178
3-bit MSE,"786,432 B","393,216 B","74,738 B","711,694 B",90.50%,80.99%,10.52x,5.26x,observed MSE 0.0348
4-bit MSE,"786,432 B","393,216 B","99,484 B","686,948 B",87.35%,74.70%,7.91x,3.95x,observed MSE 0.0098

Inner-Product-Oriented Storage (`TurboQuantProd`)
Setting,Dense float32,Dense fp16/bf16,Semafold Artifact,Saved vs float32,Smaller vs float32,Smaller vs fp16,Ratio vs float32,Ratio vs fp16,Quality Proxy
2 total bits,"786,432 B","393,216 B","50,638 B","735,794 B",93.56%,87.12%,15.53x,7.77x,mean abs dot error 0.0156
3 total bits,"786,432 B","393,216 B","75,255 B","711,177 B",90.43%,80.86%,10.45x,5.23x,mean abs dot error 0.0089
4 total bits,"786,432 B","393,216 B","99,916 B","686,516 B",87.30%,74.59%,7.87x,3.94x,mean abs dot error 0.0048

Representative Cache-Shaped Tensor Result
Workload,Shape,Dense float32,Dense fp16/bf16,Semafold Combined Artifact,Smaller vs float32,Smaller vs fp16,Ratio vs float32,Ratio vs fp16/bf16
KV block compression,"(4, 8, 256, 128)","8,388,608 B","4,194,304 B","885,734 B",89.44%,78.88%,9.47x,4.74x

K / V Split
Component,Artifact Bytes
Keys (TurboQuantProd),"459,255 B"
Values (TurboQuantMSE),"426,479 B"
Combined,"885,734 B"

Small-Block Caveat
Workload,Shape,Dense float32,Dense fp16,Semafold Combined Artifact,Smaller vs float32,Smaller vs fp16,Ratio vs float32,Ratio vs fp16
Tiny KV block,"(2, 2, 6, 16)","3,072 B","1,536 B","1,550 B",49.54%,-0.91%,1.98x,0.99x