List your company

Long-context evaluation on RULER 32K context length (test)

100Niah1 Scorefp16 Baseline
-4235077Jan 31, 2024Apr 21, 2024Jul 12, 2024Oct 1, 2024Dec 22, 2024Mar 13, 2025Jun 3, 2025

Evaluation results

MethodLinks
fp16 BaselineAvg. bit=162024.0110099.898.69468.21155.9564.537.889.6430.431.631.656.4
QuestBackbone Model=Llama-2-7B-32K-Instruct, Context Length=32K, Token Budget=1024 (3.13%)2025.0610095.8352.0887.554.17—52.3456.5126.87—36.3623.9634.3856.37
HATABackbone Model=Llama-2-7B-32K-Instruct, Context Length=32K, Token Budget=1024 (3.13%)2025.0610098.9683.3393.7578.12—65.6254.1720—43.428.1237.563.91
KVQuant-3bit-1%Avg. bit=3.332024.0199.898.895.292.861.66.447.554.4541.048.5229.33313153.65
MagicPIGBackbone Model=Llama-2-7B-32K-Instruct, Context Length=32K, Token Budget=2-3%2025.0697.9293.7554.1783.3371.88—59.3843.4916.04—51.3930.2135.4257.91
KVQuant-2bit-1%Avg. bit=2.332024.0195.486.849.873.623.4016.6522.9522.525.142426.428.436.54
DenseBackbone Model=Llama-2-7B-32K-Instruct, Context Length=32K, Token Budget=Full Attention2025.0693.7510091.6793.7581.25—66.6752.0821.04—48.6130.2136.4665.04
KIVI-2-gs32-r128Avg. bit=3.05, group size=32, residual=128-element fp162024.017685.659.672.611.4034.746.4539.68.2630.5324.827.639.78
LokiBackbone Model=Llama-2-7B-32K-Instruct, Context Length=32K, Token Budget=1024 (3.13%)2025.06252.08000—001.04—19.4414.5816.677.16
SnapKVBackbone Model=Llama-2-7B-32K-Instruct, Context Length=32K, Token Budget=1024 (3.13%)2025.06251.04000—0.2603.96—20.4922.9226.049.25
StreamingLLMBackbone Model=Llama-2-7B-32K-Instruct, Context Length=32K, Token Budget=1024 (3.13%)2025.061.045.211.046.251.04—6.254.170.21—54.8622.9227.0811.82
H2OBackbone Model=Llama-2-7B-32K-Instruct, Context Length=32K, Token Budget=1024 (3.13%)2025.060003.120—0.7800—20.4915.6216.675.15