List your company

Long-context language understanding on LongBench 1.0 (test)

61.5MultiNewsOriginal
-1.929614.537731.00547.4723Jan 13, 2024Jun 7, 2024Oct 31, 2024Mar 26, 2025Aug 19, 2025Jan 12, 2026Jun 8, 2026

Evaluation results

MethodLinks
OriginalBackbone=Llama2-7b-chat-4k, Target Context Window=4k, Result Source=LongBench (Bai et al., 2024)2025.0261.5—77.840.79.8—————31.5218.719.236.825.432.827.320.825.852.443.8
LongLora-7B-16kBackbone=Llama2-7B, Context Length=16k2024.0127.70.563.585.741.92657.654.539.7914.5534.18——————————
EntropyInferModel=Llama-3.1-8B-Instruct, Variant=w/o LD2026.0627.45—73—43.94—63.2356.47——48.9130.245.3855.255.3744.2531.6835.1825.275.9798.5
BaseModel=Llama-3.1-8B-Instruct2026.0627.3—72.5—43.66—63.0456.67——48.9429.7145.355.1355.544.331.3235.1625.316.9699.5
CritiPrefillModel=Llama-3.1-8B-Instruct2026.0627.26—72—44.18—63.2455.55——48.3227.2945.4355.4656.5443.6229.3835.0424.964.6497.5
Vicuna-v1.5-7B-16kBackbone=Llama2-7B, Version=v1.5, Context Length=16k2024.0127.215.17486.240.828.85143.536.4926.5534.28——————————
Vicuna-v1.5-13B-16kBackbone=Llama2-13B, Version=v1.5, Context Length=16k2024.0127.116.47484.927.829.844.145.638.0823.8634.92——————————
E2-LLM-Llama2-13B-16kBackbone=Llama2-13B, Context Length=16k2024.01279.873.587.940.63665.459.144.7328.5641.13——————————
H2OPrefill=100%, KV=10%, Backbone=Qwen-2.5-7B-Instruct2026.0426.81—6991.1543.97—59.2753.48——36.4922.9412.6127.3416.6315.8110.1433.5123.476.6671
E2-LLM-Llama2-13B-32kBackbone=Llama2-13B, Context Length=32k2024.0126.810.27587.840.944.563.857.544.5531.9341.74——————————
GPT-3.5-Turbo-16kContext Length=16k2024.0126.7166891.441.729.254.753.644.633.7842.19——————————
LongChat-v1.5-7B-32kBackbone=Llama2-7B, Version=v1.5, Context Length=32k2024.0126.49.963.582.334.223.25355.336.8620.4333.21——————————
EntropyInferModel=Llama-3.1-8B-Instruct2026.0626.21—73.5—44.23—62.7756.2——48.5529.8645.1854.7455.2845.8831.7430.0224.976.4998.5
Llama2-13B-chat-4kBackbone=Llama2-13B, Context Length=4k2024.0126.117.26685.236.520.351.952.837.8724.3834.87——————————
SnapKVModel=Llama-3.1-8B-Instruct2026.0626.05—67.5—42.67—62.1356.24——47.6629.6443.8954.1254.7244.8129.9928.6224.26.8299.5
AdaKVModel=Llama-3.1-8B-Instruct2026.0626—68.5—42.4—62.3356.35——47.6829.5342.9153.755.2344.8430.6228.3624.117.0199.5
Full-contextPrefill=100%, KV=100%, Backbone=Qwen-2.5-7B-Instruct2026.0425.91—73.588.9347.92—62.0366.83——49.4328.7944.7951.7758.3846.1129.2333.4624.328.83100
SnapKVPrefill=100%, KV=20%, Backbone=Qwen-2.5-7B-Instruct2026.0425.91—7288.5647.68—61.4556.84——48.2528.1344.7951.2155.8345.8228.2433.4623.858.7699.5
E2-LLM-Llama2-7B-16kBackbone=Llama2-7B, Context Length=16k2024.0125.99.668.589.238.23565.858.141.2626.738.03——————————
PI-Llama2-13B-16kBackbone=Llama2-13B, Context Length=16k2024.0125.99.172.586.527.93162.551.140.8826.3537.65——————————
SnapKVPrefill=100%, KV=10%, Backbone=Qwen-2.5-7B-Instruct2026.0425.83—70.587.6547.53—61.4556.42——47.7527.6844.7950.1255.7245.1527.1333.2523.28.1599.5
Llama2-7B-chat-4kBackbone=Llama2-7B, Context Length=4k2024.0125.80.261.577.840.719.852.443.835.1715.4520.79——————————
StreamingLLMPrefill=100%, KV=10%, Backbone=Qwen-2.5-7B-Instruct2026.0425.73—7289.4147.73—61.0656.75——46.3925.8544.7949.150.7344.9224.8333.223.27.9185
StreamingLLMPrefill=100%, KV=20%, Backbone=Qwen-2.5-7B-Instruct2026.0425.73—72.589.9647.73—62.0757.85——46.9227.9944.7949.2351.6445.1225.9433.2523.458.4785
StructKVPrefill=60%, KV=20%, Backbone=Qwen-2.5-7B-Instruct2026.0425.67—7388.6747.82—61.7861.87——48.8928.4544.7951.2557.8745.7828.7833.3424.328.83100
FullKVModel=Llama3-8B-Instruct, KV Cache Size=Full2026.0325.64—7089.8540.55—58.9652.71——48.39—37.7240.6450.1734.88—31.03—13.383.67
H2OPrefill=100%, KV=20%, Backbone=Qwen-2.5-7B-Instruct2026.0425.63—6991.0544.41—62.1754.16——37.8823.4815.7829.4319.2120.4715.1332.4721.587.1275
StructKVPrefill=60%, KV=10%, Backbone=Qwen-2.5-7B-Instruct2026.0425.53—7288.1247.26—61.6760.12——48.3928.3444.7950.6756.7645.6128.2132.1224.328.71100
E2-LLM-Llama2-7B-32kBackbone=Llama2-7B, Context Length=32k2024.0125.411.770.588.432.54064.560.941.7430.2339.18——————————
OriginalBackbone=Llama2-7b-chat-4k, Target Context Window=4k, Result Source=This paper's evaluation2025.0224.78—67.574.2440.28—56.3950.36——27.228.4813.9720.413.6216.775.4625.1712.472.33.25
ChunkLlama-16kBackbone=Llama2-7b-chat-4k, Target Context Window=16k2025.0224.78—67.574.2440.28—56.3950.36——27.228.4813.9720.413.6216.775.4625.1712.472.33.25
FastKVPrefill=60%, KV=20%, Backbone=Qwen-2.5-7B-Instruct2026.0424.67—7287.5445.87—61.1758.87——47.5528.2444.7950.3651.3444.2326.8733.3224.327.6599.5
FastKVPrefill=60%, KV=10%, Backbone=Qwen-2.5-7B-Instruct2026.0424.18—7186.2145.62—60.2257.84——46.9928.1244.7950.1250.2243.1326.2533.1224.287.2299.5
GemFilterPrefill=70%, KV=20%, Backbone=Qwen-2.5-7B-Instruct2026.0424.16—7188.5644.16—60.1246.78——45.0224.1344.2145.4549.5644.5727.1232.1923.116.2289
GemFilterPrefill=60%, KV=10%, Backbone=Qwen-2.5-7B-Instruct2026.0424.13—70.588.4343.58—59.6544.58——44.2224.0143.6844.2748.9243.8825.2331.4922.35.8887
BaseModel=Qwen2.5-7B-Instruct2026.0623.96—72—45.64—60.0767.26——48.8228.7344.2153.0557.9846.3330.0331.8623.238100
EntropyInferModel=Qwen2.5-7B-Instruct, Variant=w/o LD2026.0623.82—71—46.23—60.465.53——47.9826.8144.1452.1357.347.1630.9231.4823.66.591.5
CritiPrefillModel=Qwen2.5-7B-Instruct2026.0623.76—70—46.76—60.2765.63——47.3425.4444.251.1954.6545.9227.8131.4523.19791
FullKVModel=Qwen3-8B, KV Cache Size=Full2026.0323.69—71.6791.7942.07—70.6459.26——52.6—36.8753.6757.6744.73—33.39—11.9886.67
EntropyInferModel=Qwen2.5-7B-Instruct2026.0623.38—71—46.22—60.365.26——47.6826.8143.9851.7357.247.0930.8528.7423.186.591.5
GALI-16kBackbone=Llama2-7b-chat-4k, Target Context Window=16k2025.0223.37—6272.825.12—58.7148.51——25.86.2916.7322.265.8826.812.8223.5815.961.832.83
SnapKVModel=Qwen2.5-7B-Instruct2026.0622.75—67.5—45.03—59.7565.19——47.4328.5742.6151.3457.1644.5429.5427.0523.118100
AdaKVModel=Qwen2.5-7B-Instruct2026.0622.39—68.5—44.91—59.5565.75——47.5128.4342.251.8656.9544.5429.7726.623.118100
DapQModel=Llama3-8B-Instruct, KV Cache Size=2562026.0321.89—60.6790.4838.34—62.7855.64——46.4—32.5538.1850.6734.35—22.25—11.7883.67
PyramidKVModel=Llama3-8B-Instruct, KV Cache Size=2562026.0321.84—6089.2637.07—61.3452.51——45.7—30.8838.1150.233.88—22.54—12.7883.67
SnapKVModel=Llama3-8B-Instruct, KV Cache Size=2562026.0321.53—5789.6536.97—61.7854.92——45.61—30.8438.3949.7533.8—22.18—12.1184
H2OModel=Llama3-8B-Instruct, KV Cache Size=2562026.0321.44—45.6789.4938.28—61.4953.36——44.02—28.1136.6348.6231.5—21.87—12.1183.67
SelfExtend-16kBackbone=Llama2-7b-chat-4k, Target Context Window=16k, Result Source=LongBench (Bai et al., 2024)2025.0221.35—25.7869.581.99—60.654.33——34.6216.7721.6925.0235.2134.3430.2414.1327.325.665.83
DapQModel=Llama3-8B-Instruct, KV Cache Size=1282026.0320.63—5090.0636.87—61.8153.92——44.4—28.7637.2450.04——20.47—12.1181.67
PyramidKVModel=Llama3-8B-Instruct, KV Cache Size=1282026.0320.55—4987.6836.73—60.3652.03——43.82—28.838.2949.5231.6—20.67—12.4482
H2OModel=Llama3-8B-Instruct, KV Cache Size=1282026.0320.3—4087.2936.25—59.8153.14——42.79—25.9536.2548.6531.9—20.79—12.3383.67
SnapKVModel=Llama3-8B-Instruct, KV Cache Size=1282026.0320.08—47.6787.8235.63—61.4952.4——43.68—29.5237.849.3632.4—19.87—11.4482.33
H2OModel=Llama3-8B-Instruct, KV Cache Size=642026.0319.14—37.3386.2735.18—59.251.1——40.96—24.0230.8348.2731.7—19.37—7.7282.33
DapQModel=Qwen3-8B, KV Cache Size=2562026.0319.01—62.6791.1539.61—67.253.83——49.28—32.1450.7854.7944.47—24.16—14.1786.67
SnapKVModel=Qwen3-8B, KV Cache Size=2562026.0318.99—6391.0738.57—67.9953.77——48.92—32.449.2954.3841.4—23.79—14.6786.67
PyramidInferPrefill=60%, KV=60%, Backbone=Qwen-2.5-7B-Instruct2026.0418.94—6787.1245.62—60.2457.82——41.9525.6136.7640.1241.1241.4520.1226.3420.784.1278
DapQModel=Llama3-8B-Instruct, KV Cache Size=642026.0318.7—38.6787.3835.3—60.1949.9——41.81—25.9937.3649.1132.88—18.46—11.8977.67
H2OModel=Qwen3-8B, KV Cache Size=2562026.0318.26—46.3390.0438.98—66.1153.93——46.12—27.844.9251.3740.5—22.38—12.3386.67
PyramidKVModel=Qwen3-8B, KV Cache Size=2562026.0317.41—61.6790.9637.65—67.0150.11——47.47—30.4147.815141.37—22.36—1386.33
H2OModel=Qwen3-8B, KV Cache Size=1282026.0317.27—41.3390.2138.16—65.2252.77——44.22—26.641.3748.139.85—20.83—11.7286.67
DapQModel=Qwen3-8B, KV Cache Size=1282026.0317.27—5490.4538.5—66.2952.03——46.95—29.147.1553.8443—21.11—11.6786
SnapKVModel=Llama3-8B-Instruct, KV Cache Size=642026.0317.09—40.6786.0233.99—57.9550.91——40.78—25.0632.9247.1631.71—16.85—11.7878
SnapKVModel=Qwen3-8B, KV Cache Size=1282026.0316.64—51.3391.0737.37—65.3651.65——46.17—29.4146.4351.241.66—20.37—1186.67
PyramidKVModel=Llama3-8B-Instruct, KV Cache Size=642026.0316.34—3976.832.31—55.1947.9——38.9—22.0431.847.0131.54—15.7—10.3379.67
H2OModel=Qwen3-8B, KV Cache Size=642026.0315.23—3988.1335.98—59.4848.95——42.47—25.5538.9446.6639.27—18.55—9.6786.67
PyramidKVModel=Qwen3-8B, KV Cache Size=1282026.0315.1—48.3389.3136.92—60.8248.78——43.78—26.2240.4848.4139.46—18.99—9.6786.67
DapQModel=Qwen3-8B, KV Cache Size=642026.0313.95—4388.9736—60.3146.9——43.23—25.7843.1749.8441.28—17.16—12.6783
SelfExtend-16kBackbone=Llama2-7b-chat-4k, Target Context Window=16k, Result Source=LongBench (Jin et al., 2024)2025.0213.65—6.3167.569.88—58.2151.65——26.976.8912.6725.959.0811.257.9116.3922.792.183.21
PyramidKVModel=Qwen3-8B, KV Cache Size=642026.0312.93—44.6788.2734.63—59.7346.48——42.27—25.3240.4446.6139.2—16.25—11.3383.67
SnapKVModel=Qwen3-8B, KV Cache Size=642026.0312.38—42.6787.9335.12—57.9646.49——41.88—25.0939.8946.5839.38—15.28—11.3384.33
Dyn-NTK-16kBackbone=Llama2-7b-chat-4k, Target Context Window=16k2025.0210.9—25.9264.540.6—57.1342.16——23.573.7910.3722.3810.2626.1313.361.3224.782.294.33
YaRN-16kBackbone=Llama2-7b-chat-4k, Target Context Window=16k2025.026.31—63.545.3532.36—51.7443.91——22.223.2210.8622.143.8129.5220.1322.8413.652.22.15
NTK-16kBackbone=Llama2-7b-chat-4k, Target Context Window=16k2025.020.51—50.727.9117.08—56.3950.36——17.080.7310.3319.440.4219.476.2623.170.522.33.25