List your company

Long-context understanding on LongBench (test)

58.7Avg ScoreLlama3.1-8B
38.31643.60848.954.192Apr 25, 2024Sep 3, 2024Jan 13, 2025May 24, 2025Oct 3, 2025Feb 11, 2026Jun 23, 2026

Evaluation results

MethodLinks
Llama3.1-8BCache Bits (K,V,Index)=16, 16, 0, Backbone=Llama3.1-8B2026.0358.7—45.5—54.747.1—34.925.3————————53.87391.6——99.563.4—————56.7———————
Self-Indexing KVCacheCache Bits (K,V,Index)=16, 16, 1, Backbone=Llama3.1-8B2026.0358.4—45.4—55.447—3424.8————————53.972.591.7——99.563.2—————55.3———————
Self-Indexing KVCacheCache Bits (K,V,Index)=2, 2, 1, Backbone=Llama3.1-8B2026.0358.2—44.8—55.746—33.124.7————————54.672.592.1——99.563.2—————54.4———————
Qwen2.5-14BCache Bits (K,V,Index)=16, 16, 0, Backbone=Qwen2.5-14B2026.0356.9—46.5—6564.3—31.923.6————————548189——10033.1—————38.1———————
QuestCache Bits (K,V,Index)=16, 16, 2, Backbone=Llama3.1-8B2026.0356.5—44.9—5546.9—31.224————————51.67190.6——95.560.6—————50.5———————
Self-Indexing KVCacheCache Bits (K,V,Index)=16, 16, 1, Backbone=Qwen2.5-14B2026.0355.9—46—63.762.4—30.522.6————————53.580.588.5——97.532.6—————37.4———————
Self-Indexing KVCacheCache Bits (K,V,Index)=2, 2, 1, Backbone=Qwen2.5-14B2026.0355.7—45.8—6363.1—30.522————————53.580.588.5——96.432.3—————37.4———————
DoubleSparseCache Bits (K,V,Index)=16, 16, 2, Backbone=Llama3.1-8B2026.0354.8—43.8—54.445.9—31.724————————51.767.591.9——9750.5—————44.7———————
DoubleSparseCache Bits (K,V,Index)=16, 16, 2, Backbone=Qwen2.5-14B2026.0354.1—44.7—6162.1—30.321.9————————53.17989——94.127.5—————31.8———————
BF16Model=Llama-3.1-8B-Instruct, Bit-width=BF162025.1254—45.53—————25.3627.19———————56.3772.591.6543.62——65.1—————58.65———————
QuestCache Bits (K,V,Index)=16, 16, 2, Backbone=Qwen2.5-14B2026.0353.9—45.7—62.662.3—28.821.3————————5270.588.1——94.531.7—————35.2———————
BF16Model=Mistral-7B-Instruct-v0.3, Bit-width=BF162025.1253.84—41.58—————25.7527.82———————55.237688.5947.35——60.53—————61.68———————
Llama-3.1-8BTokens=128k, Backbone=Llama-3.1-8B2025.0553.8—45.5—54.747.1—34.9—27.5———————53.87391.643.87.556.763.4—————————————
MixKVQModel=Llama-3.1-8B-Instruct, Bit-width=C2.72025.1253.71—45.32—————25.2927.2———————55.472.591.7843.68——64.96—————57.26———————
MixKVQModel=Mistral-7B-Instruct-v0.3, Bit-width=C2.72025.1253.68—41.19—————26.2327.66———————54.527688.5946.89——60.44—————61.6———————
TailorKV-2Tokens=64(+128), Backbone=Llama-3.1-8B2025.0552.9—44.8—54.846.2—31.9—26.8———————53.970.592.343.27.754.262.1—————————————
RotateKVModel=Llama-3.1-8B-Instruct, Bit-width=KV42025.1252.89—44.32—————2527.03———————53.9271.590.8443.12——63.88—————56.39———————
KIVIModel=Mistral-7B-Instruct-v0.3, Bit-width=KV42025.1252.75—40.87—————23.926.94———————53.687488.3447.54——58.64—————60.81———————
KIVIModel=Llama-3.1-8B-Instruct, Bit-width=KV42025.1252.68—44.42—————25.2826.86———————52.172.591.4244.07——62.45—————55.04———————
TailorKV-1Tokens=64(+128), Backbone=Llama-3.1-8B2025.0552.6—43.4—55.346.5—31.3—27.2———————537091.642.57.954.861.8—————————————
SKVQModel=Llama-3.1-8B-Instruct, Bit-width=KV42025.1252.54—44.21—————25.3426.79———————53.9672.591.1442.38——62.44—————54.13———————
KVQuantModel=Mistral-7B-Instruct-v0.3, Bit-width=KV42025.1252.43—40.56—————24.7226.79———————53.417588.2146.97——58.08—————58.13———————
SKVQModel=Mistral-7B-Instruct-v0.3, Bit-width=KV42025.1252.37—40.95—————22.0426.57———————53.367689.3247.65——58.85—————56.58———————
KVQuantModel=Llama-3.1-8B-Instruct, Bit-width=KV42025.1252.3—44.03—————25.2526.87———————51.677191.0343.61——62.15—————55.13———————
ClassicSD / MagicDecTarget Model=Qwen2.5-14B-Instruct, KV Budget=None2026.0652.27—————————42.1652.9527.3671.6455.8859.29—————————————————————
PQCacheTokens=192, Backbone=Llama-3.1-8B2025.0551.7—45.6—53.845.3—29—25.9———————51.269.591.341.28.254.158.5—————————————
DustinTarget Model=Qwen2.5-14B-Instruct, KV Budget=5122026.0651.46—————————41.1152.3626.6870.8855.758.15—————————————————————
ClassicSD / MagicDecTarget Model=Qwen2.5-7B-Instruct, KV Budget=None2026.0651.46—————————40.5444.2827.7669.9253.564.11—————————————————————
KIVIModel=Mistral-7B-Instruct-v0.3, Bit-width=KV22025.1251.1—38.56—————22.4826.65———————50.477287.1347.18——57.15—————58.29———————
KIVIModel=Llama-3.1-8B-Instruct, Bit-width=KV22025.1250.72—41.02—————24.5726.79———————45.8671.591.1443.41——60.55—————51.64———————
DustinTarget Model=Qwen2.5-7B-Instruct, KV Budget=5122026.0650.4—————————40.2143.8327.4568.7553.561.45—————————————————————
DustinTarget Model=Qwen2.5-14B-Instruct, KV Budget=2562026.0650.31—————————40.2151.1926.1869.8854.4256.43—————————————————————
SKVQModel=Mistral-7B-Instruct-v0.3, Bit-width=KV22025.1249.95—32.64—————22.226.46———————46.577687.5246.09——57.9—————54.16———————
SKVQModel=Llama-3.1-8B-Instruct, Bit-width=KV22025.1249.95—35.28—————23.6126.47———————46.2772.591.2142.96——60.63—————50.63———————
KVQuantModel=Mistral-7B-Instruct-v0.3, Bit-width=KV22025.1249.94—37.88—————22.1625.89———————49.277086.5345.38——56.42—————55.94———————
SnapKVCache Bits (K,V,Index)=16, 16, 0, Backbone=Llama3.1-8B2026.0349.9—33.9—54.444.7—21.822.6————————46.44890.5——7858—————50.1———————
RotateKVModel=Mistral-7B-Instruct-v0.3, Bit-width=KV42025.1249.73—40.64—————22.925.09———————53.597688.1644.08——49.22—————47.92———————
FullKVBackbone=Qwen3-8B, KV Cache Size=Full2026.0349.46—47.76—59.2343.3736.0533.6624.0524.79———————53.3371.590.2144.43——69.39—————65.57————26.042100
DustinTarget Model=Qwen2.5-7B-Instruct, KV Budget=2562026.0649.44—————————40.4143.1727.3867.8949.2560.42—————————————————————
LOOKAHEADKVBackbone=Qwen3-8B, KV Cache Size=20482026.0349.33—48.01—59.4343.236.2132.6423.9324.93———————52.927190.2144.74——69.23—————65.01————26.761100
SpecKVBackbone=Qwen3-8B, KV Cache Size=20482026.0349.17—47—61.4343.7434.9431.3823.9824.57———————53.870.591.1144.46——69.12—————65.87————24.860100
LAQBackbone=Qwen3-8B, KV Cache Size=20482026.0349.15—47.22—57.7243.4537.431.2424.0324.51———————53.7172.590.1344.54——68.29—————64.68————24.942100
LOOKAHEADKVBackbone=Qwen3-8B, KV Cache Size=10242026.0349.04—47.23—59.343.2536.39—23.7224.61———————52.567190.2144.69——68.93—————65.22————25.360.5100
SpecKVBackbone=Qwen3-8B, KV Cache Size=10242026.0348.98—46.56—59.0443.3734.1229.3224.1823.68———————54.077190.1144.56——69.09—————66.53————24.983100
SnapKVBackbone=Qwen3-8B, KV Cache Size=20482026.0348.96—47.52—58.7342.736.0830.6423.7824.4———————53.271.590.2143.27——69.33—————65.36————25.551.1100
LOOKAHEADKVBackbone=Qwen3-8B, KV Cache Size=5122026.0348.93—46.49—59.3243.0936.9229.5623.824.01———————52.0471.590.2144.2——68.88—————65.58————25.332100
LAQBackbone=Qwen3-8B, KV Cache Size=10242026.0348.84—47.27—57.0143.5237.2629.523.8823.47———————53.4571.589.6344——67.94—————64.83————26.112100
SpecKVBackbone=Qwen3-8B, KV Cache Size=5122026.0348.79—45.65—60.5343.8535.2627.2124.0422.53———————54.3470.590.243.71——69.25—————65.85————24.223.5100
PyramidKVBackbone=Qwen3-8B, KV Cache Size=20482026.0348.78—46.69—58.4142.936.6129.4123.6124.18———————53.2171.590.0542.9——69.28—————65.21————25.471.1100
LAQBackbone=Qwen3-8B, KV Cache Size=5122026.0348.77—46.98—57.3143.3537.6426.9323.6722.19———————53.772.588.9643.81——68.23—————64.71————27.343100
SnapKVBackbone=Qwen3-8B, KV Cache Size=10242026.0348.45—46.13—58.5242.6636.8928.3923.6123.33———————52.486989.5543.13——69.05—————66.27————24.262100
LOOKAHEADKVBackbone=Qwen3-8B, KV Cache Size=2562026.0348.31—45.4—58.4744.0536.1327.7723.7122.88———————52.686989.0543.32——67.83—————64.71————25.882100
TriAttentionModel=Qwen3-8B, KV budget=50%2026.0448.128.14351.460.244.936.932.923.824.3————————6990.339.9—9165—————61.3—————7—
Sketch&WalkBackbone=Llama3.1-8B-Instruct2026.024830.6944.44—57.4947.1133.1632.3925.4725.55———————55.7370.591.443.398.759954.94—————47.7850.6——————
PyramidKVBackbone=Qwen3-8B, KV Cache Size=10242026.0348—42.89—58.8642.3235.4727.3223.0722.72———————53.017189.9542.56——68.81—————64.25————23.772100
DenseBackbone=Llama3.1-8B-Instruct2026.0247.9530.0544.67—55.445.6229.4134.7725.1426.9———————55.977391.1643.24109955.13—————47.7950.48——————
LAQBackbone=Qwen3-8B, KV Cache Size=2562026.0347.95—45.44—5743.5336.6224.2223.3820.38———————53.847089.0542.47——68.17—————64.03————263100
KVQuantModel=Llama-3.1-8B-Instruct, Bit-width=KV22025.1247.9—36.51—————22.9425.82———————41.856586.2741.94——59.88—————50.87———————
AdamasBackbone=Llama3.1-8B-Instruct2026.0247.8430.0943.84—56.0745.8131.234.7225.0526.89———————56.27391.2943.1489954.2—————46.9650.49——————
SnapKVBackbone=Qwen3-8B, KV Cache Size=5122026.0347.64—43.72—58.3742.3634.0425.0322.5521.66———————51.966989.5342.06——69.37—————64.96————24.633100
DustinTarget Model=Qwen2.5-14B-Instruct, KV Budget=1282026.0647.59—————————40.5949.7525.5769.4938.8653.72—————————————————————
ClassicSD / MagicDecTarget Model=Qwen2.5-32B-Instruct, KV Budget=None2026.0647.52—————————42.4354.5327.3767.575642.78—————————————————————
LOOKAHEADKVBackbone=Qwen3-8B, KV Cache Size=1282026.0347.46—44.3—58.7842.7935.8925.2922.9521.13———————53.2466.588.9541.64——65.95—————62.88————26.063.599.5
SpecKVBackbone=Qwen3-8B, KV Cache Size=2562026.0347.42—41.09—59.8542.4234.524.5323.6421.25———————53.896888.1343.12——68.39—————64.4————22.583100
Full Attn.Model=Qwen3-8B, KV budget=Full2026.0447.228.843.855.362.848.935.533.524.724.7————————40.590.540.3—91.864.9—————60—————9—
Yi-9BTokens=200k, Backbone=Yi-9B2025.0547—38.4—52.736.7—31—26.7———————34.9779014.92.567.471.9—————————————
QuestBackbone=Llama3.1-8B-Instruct2026.0246.9228.7942.21—55.9346.9530.8934.824.1927.32———————55.986989.942.648.039949.96—————45.1249.51——————
NSA2025.0246.9—43.2—43.735.6—30.7—————————50.3————90.523.262.40.34155——————————
DustinTarget Model=Qwen2.5-32B-Instruct, KV Budget=5122026.0646.85—————————41.3754.526.666.7556.2541.73—————————————————————
LAQBackbone=Qwen3-8B, KV Cache Size=1282026.0346.52—42.15—57.8942.8436.7421.3322.2518.34———————53.5564.589.5540.93——66.74—————61.7————22.743100
PyramidKVBackbone=Qwen3-8B, KV Cache Size=5122026.0346.44—40.52—57.5740.8932.8523.8521.9219.7———————51.4368.589.5540.89——67.47—————61.73————23.123100
DustinTarget Model=Qwen2.5-7B-Instruct, KV Budget=1282026.0646.29—————————39.8742.7526.7667.8435.555.04—————————————————————
QuestTokens=192, Backbone=Llama-3.1-8B2025.0546.2—35.9—52.841—17.7—23.8———————44.26386368.443.652.3—————————————
DustinTarget Model=Qwen2.5-32B-Instruct, KV Budget=2562026.0645.72—————————41.2954.125.7164.8753.540.53—————————————————————
SpecKVBackbone=Qwen3-8B, KV Cache Size=1282026.0345.62—37.14—56.7742.2431.8221.3322.8619.04———————53.586088.3141.5——66.82—————61.96————23.033.5100
Ada-KV+SnapKVModel=Qwen3-8B, KV budget=50%2026.0445.62736.84559.647.534.131.82323.6————————4690.140.9—90.865.6—————59.6—————8—
SnapKVBackbone=Qwen3-8B, KV Cache Size=2562026.0345.45—38.32—57.3640.6732.8221.5121.8918.97———————51.0459.589.4641.06——67.62—————61.88————23.032100
SnapKVTokens=192, Backbone=Llama-3.1-8B2025.0545.2—25.7—52.643.7—20—20.5———————44.7418939.6848.757—————————————
SnapKVModel=Qwen3-8B, KV budget=50%2026.0445.226.93745.559.246.333.132.22323.4————————40.589.941—91.166.4—————59.9—————7.5—
GPT-4-Turboparadigm=Zero-shot, truncation=Middle truncation strategy, decoding=Greedy decoding, context_limit=32K2024.0444.73350.752.768.564.349.133.925.424.9———————————————————————————
TailorKV-1Tokens=64(+128), Backbone=Yi-9B2025.0544.7—37.7—52.835.8—29.8—24.9———————38.27689.515.23.564.368.4—————————————
SnapKVCache Bits (K,V,Index)=16, 16, 0, Backbone=Qwen2.5-14B2026.0344.3—29.4—60.154—19.319.8————————47.654.586——6124.9—————31.1———————
TailorKV-2Tokens=64(+128), Backbone=Yi-9B2025.0544—37.6—52.634.4—29.4—25.3———————33.67689.315363.568.5—————————————
Full Attention2025.0243.7—40.9—3530.5—32.4—————————51.2————8316.362.30.29456——————————
LOOKAHEADKVBackbone=Qwen3-8B, KV Cache Size=642026.0343.53—43.13—59.0142.534.3422.6621.6118.49———————51.8564.588.7539.48——62.78—————56.07————22.111.567.75
DustinTarget Model=Qwen2.5-32B-Instruct, KV Budget=1282026.0643.28—————————41.1253.724.7662.8137.9239.77—————————————————————
PyramidKVBackbone=Qwen3-8B, KV Cache Size=2562026.0343.05—34.87—55.6837.8926.6720.4320.9217.43———————47.4458.585.238.98——65.51—————57.32————18.473.5100
PyramidKVModel=Qwen3-8B, KV budget=50%2026.0442.725.930.439.152.239.929.729.922.221.9————————33.59040.7—93.465.7—————60.2—————8—
Exact-Top2025.0242.3—39.7—32.128.8—31.6—————————50.2————8115.660.50.29154.8——————————
SnapKVBackbone=Qwen3-8B, KV Cache Size=1282026.0342.13—32.65—54.8138.9526.5917.6620.8316.04———————45.9949.587.138.9——64.62—————58.29————19.143.599.5
PQCacheTokens=192, Backbone=Yi-9B2025.0542—36.9—47.635.5—19.3—19.2———————27.97489.612.24.66266.8—————————————
StreamingLLMBackbone=Qwen3-8B, KV Cache Size=20482026.0341.47—38.54—47.9636.7825.5328.6921.4724.11———————38.026590.342.85——68.37—————64.63————21.73148.5
StreamLLMTokens=192, Backbone=Llama-3.1-8B2025.0540.6—21.4—46.538.9—17.9—18———————31.34075.734.4845.750.7—————————————
QuestTarget Model=Qwen2.5-32B-Instruct, KV Budget=5122026.0640.41—————————31.336.8824.1659.5554.9740—————————————————————
FILM-7Bparadigm=Zero-shot, truncation=Middle truncation strategy, decoding=Greedy decoding, context_limit=32K2024.0439.926.942.25662.1473933.825.126.9———————————————————————————
QuestTarget Model=Qwen2.5-14B-Instruct, KV Budget=5122026.0639.83—————————29.2730.5125.0460.6649.7343.67—————————————————————
StreamingLLMModel=Qwen3-8B, KV budget=50%2026.0439.424.130.531.246.541.621.630.821.823.9————————4385.438.2—55.264.7—————61.2—————10—
Quest2025.0239.2—36.5—29.524.5—29.3—————————49.5————79.213.556.10.25747.8——————————
StreamingLLMBackbone=Qwen3-8B, KV Cache Size=10242026.0339.19—32.82—45.9434.3823.3425.7320.2523.5———————31.446288.7141.18——68.39—————63.65————21.250.544
QuestTokens=192, Backbone=Yi-9B2025.0539.1—32.4—44.231.5—14.3—16.5———————267386.213.43.955.663.7—————————————