List your company

Prefill KV-cache memory measurement on TULU-3 (dev)

0.109Active KV-cache Memory (GiB)SPEED-24
-0.527283.767618.062512.35739May 7, 2026

Evaluation results

MethodLinks
SPEED-24Prompt=1K, K=242026.050.10922.7
SwiftKV-24Prompt=1K, K=242026.050.12312.8
Full-AttnPrompt=1K2026.050.1410
POP-24Prompt=1K, K=242026.050.1410
SPEED-24Prompt=2K, K=242026.050.20323.7
SwiftKV-24Prompt=2K, K=242026.050.23212.8
Full-AttnPrompt=2K2026.050.2660
POP-24Prompt=2K, K=242026.050.2660
SPEED-24Prompt=4K, K=242026.050.39124.2
SwiftKV-24Prompt=4K, K=242026.050.45112.6
Full-AttnPrompt=4K2026.050.5160
POP-24Prompt=4K, K=242026.050.5160
SPEED-24Prompt=8K, K=242026.050.76624.6
SwiftKV-24Prompt=8K, K=242026.050.88912.5
Full-AttnPrompt=8K2026.051.0160
POP-24Prompt=8K, K=242026.051.0160
SPEED-24Prompt=16K, K=242026.051.51624.8
SwiftKV-24Prompt=16K, K=242026.051.76412.5
Full-AttnPrompt=16K2026.052.0160
POP-24Prompt=16K, K=242026.052.0160
SPEED-24Prompt=32K, K=242026.053.01624.9
SwiftKV-24Prompt=32K, K=242026.053.51412.5
Full-AttnPrompt=32K2026.054.0160
POP-24Prompt=32K, K=242026.054.0160
SPEED-24Prompt=64K, K=242026.056.01625
SwiftKV-24Prompt=64K, K=242026.057.01412.5
Full-AttnPrompt=64K2026.058.0160
POP-24Prompt=64K, K=242026.058.0160
SPEED-24Prompt=128K, K=242026.0512.01625
SwiftKV-24Prompt=128K, K=242026.0514.01412.5
Full-AttnPrompt=128K2026.0516.0160
POP-24Prompt=128K, K=242026.0516.0160