List your company

Long-context Evaluation on LongBench

55.26Average ScoreIndexMem
2.043215.859129.67543.4909Feb 6, 2025May 2, 2025Jul 26, 2025Oct 19, 2025Jan 12, 2026Apr 7, 2026Jul 1, 2026

Evaluation results

MethodLinks
IndexMemCompression ratio=50%2026.0555.26———————
IndexMemCompression ratio=10%2026.0553.48———————
IndexMemCompression ratio=25%2026.0553.18———————
FullKVBackbone=Qwen3-8B, KV Cache Size=Full2026.0352.6———————
FullKVBackbone=Qwen2.5-7B-Instruct, KV Cache Size=Full2026.0350.67———————
DapQBackbone=Qwen3-8B, KV Cache Size=2562026.0349.28———————
SnapKVBackbone=Qwen3-8B, KV Cache Size=2562026.0348.92———————
FullKVBackbone=Llama3-8B-Instruct, KV Cache Size=Full2026.0348.39———————
PyramidKVBackbone=Qwen3-8B, KV Cache Size=2562026.0347.47———————
DapQBackbone=Qwen3-8B, KV Cache Size=1282026.0346.95———————
DapQBackbone=Qwen2.5-7B-Instruct, KV Cache Size=2562026.0346.45———————
DapQBackbone=Llama3-8B-Instruct, KV Cache Size=2562026.0346.4———————
SnapKVBackbone=Qwen2.5-7B-Instruct, KV Cache Size=2562026.0346.19———————
SnapKVBackbone=Qwen3-8B, KV Cache Size=1282026.0346.17———————
H2OBackbone=Qwen3-8B, KV Cache Size=2562026.0346.12———————
PyramidKVBackbone=Llama3-8B-Instruct, KV Cache Size=2562026.0345.7———————
SnapKVBackbone=Llama3-8B-Instruct, KV Cache Size=2562026.0345.61———————
PyramidKVBackbone=Qwen2.5-7B-Instruct, KV Cache Size=2562026.0344.82———————
DapQBackbone=Llama3-8B-Instruct, KV Cache Size=1282026.0344.4———————
H2OBackbone=Qwen2.5-7B-Instruct, KV Cache Size=2562026.0344.31———————
H2OBackbone=Qwen3-8B, KV Cache Size=1282026.0344.22———————
H2OBackbone=Llama3-8B-Instruct, KV Cache Size=2562026.0344.02———————
Expected AttentionCompression ratio=25%2026.0543.88———————
PyramidKVBackbone=Llama3-8B-Instruct, KV Cache Size=1282026.0343.82———————
DapQBackbone=Qwen2.5-7B-Instruct, KV Cache Size=1282026.0343.82———————
TOVACompression ratio=10%2026.0543.81———————
PyramidKVBackbone=Qwen3-8B, KV Cache Size=1282026.0343.78———————
Expected AttentionCompression ratio=10%2026.0543.78———————
SnapKVBackbone=Llama3-8B-Instruct, KV Cache Size=1282026.0343.68———————
PyramidKVCompression ratio=10%2026.0543.63———————
FP16BIT=162026.0743.5922.947.3954.2222.630.3188.0439.67
SnapKVCompression ratio=10%2026.0543.55———————
HeadKV w/ CriticalKVBackbone=Mistral-7B, Cache Budget=20%2025.0243.46———————
SnapKVCompression ratio=25%2026.0543.37———————
DapQBackbone=Qwen3-8B, KV Cache Size=642026.0343.23———————
TOVACompression ratio=25%2026.0543.08———————
SnapKVBackbone=Qwen2.5-7B-Instruct, KV Cache Size=1282026.0343.02———————
AdaKV w/ CriticalKVBackbone=Mistral-7B, Cache Budget=20%2025.0242.85———————
H2OBackbone=Llama3-8B-Instruct, KV Cache Size=1282026.0342.79———————
H2OBackbone=Qwen3-8B, KV Cache Size=642026.0342.47———————
PyramidKVCompression ratio=50%2026.0542.35———————
PyramidKVBackbone=Qwen3-8B, KV Cache Size=642026.0342.27———————
H2OBackbone=Qwen2.5-7B-Instruct, KV Cache Size=1282026.0342.22———————
IndexMemCompression ratio=75%2026.0542.18———————
TOVACompression ratio=50%2026.0542.1———————
Expected AttentionCompression ratio=50%2026.0542.09———————
PyramidKVBackbone=Qwen2.5-7B-Instruct, KV Cache Size=1282026.0342.03———————
SnapKVBackbone=Qwen3-8B, KV Cache Size=642026.0341.88———————
DapQBackbone=Llama3-8B-Instruct, KV Cache Size=642026.0341.81———————
SnapKV w/ CriticalKVBackbone=Mistral-7B, Cache Budget=20%2025.0241.77———————
PyramidKVCompression ratio=25%2026.0541.57———————
SnapKVCompression ratio=50%2026.0541.44———————
DapQBackbone=Qwen2.5-7B-Instruct, KV Cache Size=642026.0341.33———————
H2OBackbone=Llama3-8B-Instruct, KV Cache Size=642026.0340.96———————
SnapKVBackbone=Llama3-8B-Instruct, KV Cache Size=642026.0340.78———————
SnapKV w/ OBCache (Joint)Backbone=Mistral-7B, Cache Budget=20%2025.0240.49———————
SnapKVBackbone=Mistral-7B, Cache Budget=20%2025.0240.11———————
SnapKVBackbone=Qwen2.5-7B-Instruct, KV Cache Size=642026.0340.09———————
H2OBackbone=Qwen2.5-7B-Instruct, KV Cache Size=642026.0339.98———————
GSRQBIT=2.002026.0739.3921.9644.8848.2422.6724.4275.338.27
SnapKVCompression ratio=75%2026.0539.22———————
PyramidKVCompression ratio=75%2026.0539.17———————
TOVACompression ratio=75%2026.0539.05———————
PyramidKVBackbone=Llama3-8B-Instruct, KV Cache Size=642026.0338.9———————
VQLLMBIT=2.002026.0738.4723.3842.246.4622.1421.1477.1236.85
PyramidKVBackbone=Qwen2.5-7B-Instruct, KV Cache Size=642026.0338.45———————
CompilerKVModel=LLaMA-3-8B-Instruct, KV cache size=1282026.0237.97———————
CompilerKVModel=Mistral-7B-Instruct-v0.2, KV cache size=1282026.0237.63———————
IndexMemCompression ratio=90%2026.0536.51———————
CompilerKVModel=InternLM-2.5-Chat-1M, KV cache size=1282026.0236.5———————
Expected AttentionCompression ratio=75%2026.0536.33———————
CompilerKVModel=Qwen2-7B-Instruct, KV cache size=1282026.0236.13———————
TOVACompression ratio=90%2026.0532.56———————
LOOKAHEADKVBackbone=Llama3.1-8B, KV Cache Size=2562026.0331.96———————
FullKVBackbone=Llama3.1-8B2026.0331.63———————
LOOKAHEADKVBackbone=Llama3.1-8B, KV Cache Size=5122026.0331.39———————
LOOKAHEADKVBackbone=Llama3.1-8B, KV Cache Size=1282026.0331.32———————
LOOKAHEADKVBackbone=Llama3.1-8B, KV Cache Size=10242026.0331.14———————
LOOKAHEADKVBackbone=Llama3.1-8B, KV Cache Size=20482026.0331.01———————
LOOKAHEADKVBackbone=Llama3.1-8B, KV Cache Size=642026.0330.62———————
PyramidKVCompression ratio=90%2026.0529.68———————
SnapKVCompression ratio=90%2026.0529.4———————
Expected AttentionCompression ratio=90%2026.0528.6———————
LOOKAHEADKVBackbone=Qwen3-8B, KV Cache Size=20482026.0326.76———————
LOOKAHEADKVBackbone=Qwen3-8B, KV Cache Size=1282026.0326.06———————
FullKVBackbone=Qwen3-8B2026.0326.04———————
LOOKAHEADKVBackbone=Qwen3-8B, KV Cache Size=2562026.0325.88———————
LOOKAHEADKVBackbone=Qwen3-8B, KV Cache Size=10242026.0325.36———————
LOOKAHEADKVBackbone=Qwen3-8B, KV Cache Size=5122026.0325.33———————
LOOKAHEADKVBackbone=Qwen3-8B, KV Cache Size=642026.0322.11———————
GSRQBIT=1.002026.0720.2118.5630.2840.7721.4811.888.1710.35
GSRQBIT=0.752026.0714.9612.3917.4229.721.115.467.918.83
VQLLMBIT=1.002026.0712.236.9915.7324.4118.353.7311.434.96
MambaTraining=compute-matched, Training dataset=FineWeb-edu2026.054.96———————
TransformerTraining=compute-matched, Training dataset=FineWeb-edu2026.054.44———————
SRMTraining=compute-matched, Training dataset=FineWeb-edu2026.054.09———————