List your company

Long-context language understanding on LongBench (test)

51.87Average ScoreFull-context
33.950838.602943.25547.9071Feb 3, 2025Apr 16, 2025Jun 27, 2025Sep 8, 2025Nov 19, 2025Jan 30, 2026Apr 13, 2026

Evaluation results

MethodLinks
Full-contextModel=Ministral-8B-Instruct, Prefill=100%, KV=100%, Execution strategy=Full-context2025.0251.87————————————————41.5849.4527.7270.8367.1354.5—————————
SnapKVModel=Ministral-8B-Instruct, Prefill=100%, KV=20%, Execution strategy=Decoding-Only2025.0251.35————————————————40.6349.5925.6770.6467.0754.5—————————
FastKVModel=Ministral-8B-Instruct, Prefill=60%, KV=20%, Execution strategy=Prefill-Aware2025.0251.07————————————————41.0849.225.0271.4164.6855—————————
SnapKVModel=Ministral-8B-Instruct, Prefill=100%, KV=10%, Execution strategy=Decoding-Only2025.0250.68————————————————39.9749.4323.9270.0365.7555—————————
FastKVModel=Ministral-8B-Instruct, Prefill=60%, KV=10%, Execution strategy=Prefill-Aware2025.0250.4————————————————39.7149.6323.3870.716455—————————
Full-contextModel=LLaMA-3.1-8B-Instruct, Prefill=100%, KV=100%, Execution strategy=Full-context2025.0250.19————————————————43.5844.6529.2269.4860.0154.21—————————
AdmTreeBackbone=Qwen-2-7B, Method category=Compression-based Methods, Latency=7.0, Speedup=3.4x2025.1249.7————————————————41.645.930.269.966.8——————————
SnapKVModel=LLaMA-3.1-8B-Instruct, Prefill=100%, KV=20%, Execution strategy=Decoding-Only2025.0249.43————————————————43.2743.9226.7568.2160.7353.73—————————
FastKVModel=LLaMA-3.1-8B-Instruct, Prefill=60%, KV=20%, Execution strategy=Prefill-Aware2025.0249.07————————————————42.9343.7426.0369.5168.6453.56—————————
FullBackbone=Llama-3.1-8B, Budget=Full2026.0149.0631.0645.4353.7855.0447.1431.2934.8725.3327.4972.591.2543.81699.563.3656.65———————————————
FullBackbone=Qwen2.5-7B, Budget=Full2026.0148.8729.0543.3452.5257.5947.0530.2431.7823.6423.9672.589.4745.618.510059.6167.12———————————————
TAPPABackbone=Llama-3.1-8B, Budget=20482026.0148.7330.745.6953.0655.4946.6830.9430.5424.6527.127191.6543699.564.9358.8———————————————
SnapKVModel=LLaMA-3.1-8B-Instruct, Prefill=100%, KV=10%, Execution strategy=Decoding-Only2025.0248.73————————————————41.944.1125.3468.1759.5353.72—————————
PyramidKVBackbone=Llama-3.1-8B, Budget=20482026.0148.5131.1345.0653.855.7846.5930.8930.2524.8227.357191.6542.62699.563.2756.44———————————————
Full-KVModel=Qwen2.5-32B, Compression ratio=0%2026.0248.51————————————————42.9154.1527.3368.9142.3555.75—————————
FastKVModel=LLaMA-3.1-8B-Instruct, Prefill=60%, KV=10%, Execution strategy=Prefill-Aware2025.0248.47————————————————41.9543.7324.7869.4557.4553.43—————————
TAPPABackbone=Qwen2.5-7B, Budget=20482026.0148.4629.1844.0352.2457.3645.7730.1929.1423.3123.626989.4744.99810060.9568.06———————————————
TAPPABackbone=Llama-3.1-8B, Budget=10242026.0148.4330.7744.9452.1455.4346.9931.1628.7224.926.6569.591.9542.38699.564.9958.84———————————————
CAKEBackbone=Llama-3.1-8B, Budget=20482026.0148.4330.7945.8353.5755.546.630.4731.1224.6727.1670.591.4843.48699.563.2356.64———————————————
SnapKVBackbone=Llama-3.1-8B, Budget=20482026.0148.3230.9945.0653.1555.2546.5630.7830.2424.6327.3270.591.4842.37699.563.2856.86———————————————
CAKEBackbone=Qwen2.5-7B, Budget=20482026.0148.3129.0843.3551.9257.245.7730.2629.3523.4623.596989.3745.37810059.3567.88———————————————
LU-KVModel=Qwen2.5-32B, Eviction Metric=KeyDiff Metric (π2), Compression ratio=80%2026.0248.26————————————————41.5854.2326.6167.8646.9153.75—————————
PyramidKVBackbone=Qwen2.5-7B, Budget=20482026.0148.1728.3943.3651.8356.7545.630.526.923.0323.387188.2245.01810061.3667.37———————————————
CAKEBackbone=Llama-3.1-8B, Budget=10242026.0148.1330.8844.9552.3855.4946.9930.8228.6824.9126.396991.9442.6699.562.6556.89———————————————
SnapKVBackbone=Qwen2.5-7B, Budget=20482026.0148.1229.1141.5352.0557.1746.2630.6929.4923.2323.6471.589.1745.49810060.9267.94———————————————
SnapKVBackbone=Llama-3.1-8B, Budget=10242026.0147.9530.9544.7452.5855.0946.8330.3727.8724.5725.9968.592.0342.66.599.56356.5———————————————
LU-KVModel=Qwen2.5-32B, Eviction Metric=SnapKV Metric (π1), Compression ratio=80%2026.0247.95————————————————39.8453.5526.1967.3248.4654.75—————————
TAPPABackbone=Qwen2.5-7B, Budget=10242026.0147.7229.6443.5951.5356.945.8629.4326.6422.572365.589.4845.24810060.0466.06———————————————
CAKEBackbone=Qwen2.5-7B, Budget=10242026.0147.729.4742.7152.1256.1146.4129.1326.8623.1222.7267.589.2345.468.510059.1164.79———————————————
PyramidKVBackbone=Llama-3.1-8B, Budget=10242026.0147.6930.5443.6452.7355.2946.2931.2827.5324.5266892.0941.756.0599.562.3555.44———————————————
Original LLM-FTBackbone=Qwen-2-7B, Latency=23.52025.1247.4————————————————4140.626.868.566.1——————————
H2OBackbone=Llama-3.1-8B, Budget=20482026.0147.329.6539.5348.6454.2346.529.2829.9723.6827.2168.591.4843.066.1199.563.0656.91———————————————
Full-KVModel=Mistral-7B-v0.3, Compression ratio=0%2026.0247.3————————————————38.3637.8328.8670.8663.2651.25—————————
SnapKVBackbone=Qwen2.5-7B, Budget=10242026.0147.2729.2441.6150.9357.645.529.3925.6323.0622.2665.588.9244.658.510058.1665.3———————————————
TAPPABackbone=Llama-3.1-8B, Budget=5122026.0147.2129.4742.6651.6354.5346.6430.8125.4824.5724.7162.592.3542.426.2599.564.5657.35———————————————
BeaconBackbone=Qwen-2-7B, Method category=Compression-based Methods, Latency=7.3, Speedup=3.2x2025.1247.2————————————————40.540.326.868.466.4——————————
CAKEBackbone=Llama-3.1-8B, Budget=5122026.0147.1931.8242.9951.6554.3746.8930.7326.3624.9425.2763.591.5442.526.3399.562.354.3———————————————
PyramidInferModel=Ministral-8B-Instruct, Prefill=60%, KV=60%, Execution strategy=Prefill-Aware2025.0247.01————————————————34.2937.3424.1668.1965.8352.25—————————
H2OBackbone=Qwen2.5-7B, Budget=20482026.0146.9528.5540.247.4553.4944.442728.9322.6623.7963.588.546.08810061.0667.5———————————————
H2OModel=Ministral-8B-Instruct, Prefill=100%, KV=20%, Execution strategy=Decoding-Only2025.0246.95————————————————38.9344.2726.3869.363.5839.25—————————
SnapKVBackbone=Llama-3.1-8B, Budget=5122026.0146.9230.7642.0352.1354.1546.1430.5124.9824.2424.656492.0542.046.0899.562.6254.9———————————————
PyramidKVBackbone=Llama-3.1-8B, Budget=5122026.0146.5930.4742.1552.1754.6745.2530.62524.3324.5162.591.2441.675.9599.561.5853.89———————————————
GemFilterModel=Ministral-8B-Instruct, Prefill=70%, KV=20%, Execution strategy=Prefill-Aware2025.0246.37————————————————40.8454.0625.3168.238.5651.25—————————
LU-KVModel=Mistral-7B-v0.3, Eviction Metric=KeyDiff Metric (π2), Compression ratio=80%2026.0246.21————————————————39.838.0928.2268.3256.0452.02—————————
H2OBackbone=Llama-3.1-8B, Budget=10242026.0146.1129.5736.1545.9454.4344.8129.0427.6423.3126.476291.4343.146.369962.2455.74———————————————
H2OModel=Ministral-8B-Instruct, Prefill=100%, KV=10%, Execution strategy=Decoding-Only2025.0245.94————————————————37.4743.8825.9668.6460.9538.75—————————
PyramidKVBackbone=Qwen2.5-7B, Budget=10242026.0145.8529.3438.650.1755.6745.1227.8223.2622.1620.5562.586.8543.268.510057.7661.99———————————————
LU-KVModel=Mistral-7B-v0.3, Eviction Metric=SnapKV Metric (π1), Compression ratio=80%2026.0245.79————————————————37.1636.5927.9769.8157.6951.35—————————
TAPPABackbone=Qwen2.5-7B, Budget=5122026.0145.7828.9739.450.4655.4844.4728.0223.9922.8720.725587.0244.66810059.0464.39———————————————
Original LLMBackbone=Qwen-2-7B, Latency=23.5, Speedup=1.0x2025.1245.7————————————————38.837.526.770.160.3——————————
PyramidKVBackbone=Qwen2.5-7B, Budget=5122026.0145.5827.3338.0450.3855.7344.2827.1222.2421.8619.546686.3643.6989957.5962.09———————————————
CAKEBackbone=Qwen2.5-7B, Budget=5122026.0145.5628.9739.4650.454.844.728.0223.922.3520.745586.9144.92899.557.0664.26———————————————
SnapKVBackbone=Qwen2.5-7B, Budget=5122026.0145.5128.9440.750.455.844.2127.8324.4222.7421.0766.586.5644.14899.559.1764.22———————————————
SnapKVBackbone=Qwen-2-7B, Method category=Compression-based Methods, Latency=7.8, Speedup=3.0x2025.1245————————————————38.737.626.267.160.3——————————
DensePer-Query Budget=-, Backbone=Trado-8B-Instruct2026.0444.8619.0417.75—49.45——————84.79———————————————————53.29
H2OBackbone=Llama-3.1-8B, Budget=5122026.0144.3927.7629.0144.7552.7844.3129.2224.7123.1124.5654.591.3842.16.369962.354.33———————————————
StreamingLLMBackbone=Llama-3.1-8B, Budget=20482026.0144.3927.436.9137.8549.2344.6624.3128.5721.6727.1267.590.9842.496.128763.0655.32———————————————
StreamingLLMModel=LLaMA-3.1-8B-Instruct, Prefill=100%, KV=20%, Execution strategy=Decoding-Only2025.0244.2————————————————30.4338.8924.1965.4159.4446.83—————————
H2OBackbone=Qwen2.5-7B, Budget=10242026.0144.1426.4534.9440.4948.6342.0222.2725.6720.922.415987.8345.078.598.4859.7763.88———————————————
AdmTreeBackbone=Llama-2-7B, Method category=Compression-based Methods, Latency=7.8, Speedup=3.3x2025.1244.1————————————————36.536.326.965.560.9——————————
H2OBackbone=Qwen2.5-7B, Budget=5122026.0144.0726.8334.1741.4350.841.8322.8225.5721.3522.0360.584.6745.86895.559.1164.66———————————————
LoSAPer-Query Budget=1024, Backbone=Trado-8B-Instruct2026.0442.8818.8915.78—48.45——————82.32———————————————————48.94
StreamingLLMModel=LLaMA-3.1-8B-Instruct, Prefill=100%, KV=10%, Execution strategy=Decoding-Only2025.0242.85————————————————28.6737.0122.2863.1958.0147.96—————————
GemFilterModel=Ministral-8B-Instruct, Prefill=60%, KV=10%, Execution strategy=Prefill-Aware2025.0242.82————————————————37.9255.2223.365.0530.8846—————————
StreamingLLMBackbone=Llama-3.1-8B, Budget=10242026.0142.7326.6430.7735.5947.3142.0324.1725.8121.3125.6663.588.8442.766.58861.3153.47———————————————
GemFilterModel=LLaMA-3.1-8B-Instruct, Prefill=61%, KV=20%, Execution strategy=Prefill-Aware2025.0242.39————————————————34.6340.7724.3966.3936.5651.59—————————
AdaModel=Qwen2.5-32B, Eviction Metric=KeyDiff Metric (π2), Compression ratio=80%2026.0242.32————————————————32.1646.4423.6866.3336.7148.92—————————
LoSAPer-Query Budget=512, Backbone=Trado-8B-Instruct2026.0442.218.3913.3—44.19——————84.97———————————————————50.14
LoSAPer-Query Budget=256, Backbone=Trado-8B-Instruct2026.0442.1419.4217.11—44.53——————81.82———————————————————47.81
LoSAPer-Query Budget=128, Backbone=Trado-8B-Instruct2026.0441.9717.1915.58—48.27——————83.79———————————————————45
AdaModel=Mistral-7B-v0.3, Eviction Metric=SnapKV Metric (π1), Compression ratio=80%2026.0241.89————————————————26.2131.8824.4266.5262.0549.5—————————
StreamingLLMBackbone=Llama-3.1-8B, Budget=5122026.0141.7525.6427.4833.347.3640.0624.823.1620.822.8557.587.642.086.59760.5151.28———————————————
AdaModel=Qwen2.5-32B, Eviction Metric=SnapKV Metric (π1), Compression ratio=80%2026.0241.58————————————————25.7543.623.3465.744.9750.13—————————
UniformModel=Qwen2.5-32B, Eviction Metric=SnapKV Metric (π1), Compression ratio=80%2026.0241.21————————————————25.0244.523.3364.4544.9948.75—————————
StreamingLLMModel=Ministral-8B-Instruct, Prefill=100%, KV=10%, Execution strategy=Decoding-Only2025.0241.21————————————————27.7435.872165.4462.734.5—————————
PyramidModel=Mistral-7B-v0.3, Eviction Metric=SnapKV Metric (π1), Compression ratio=80%2026.0240.94————————————————25.3631.9523.9863.3561.848.75—————————
UniformModel=Mistral-7B-v0.3, Eviction Metric=SnapKV Metric (π1), Compression ratio=80%2026.0240.67————————————————24.731.7824.1563.9560.9847.5—————————
AdaModel=Mistral-7B-v0.3, Eviction Metric=KeyDiff Metric (π2), Compression ratio=80%2026.0240.54————————————————31.9433.7825.2359.657.7140.79—————————
Dense FlashAttention-2Sparsity=All2025.1240.24——————————————————————1————————
BeaconBackbone=Llama-2-7B, Method category=Compression-based Methods, Latency=8.0, Speedup=3.2x2025.1240.1————————————————34.927.52561.457.8——————————
BSFASparsity=k=1282025.1240.03——————————————————————1.08————————
BSFASparsity=k=962025.1239.88——————————————————————1.1————————
Original LLM-FTBackbone=Llama-2-7B, Latency=25.62025.1239.8————————————————34.827.523.261.857.8——————————
H2OModel=LLaMA-3.1-8B-Instruct, Prefill=100%, KV=20%, Execution strategy=Decoding-Only2025.0239.79————————————————30.6619.6926.056858.0536.28—————————
BSFASparsity=k=642025.1239.78——————————————————————1.13————————
BSFASparsity=k=322025.1239.39——————————————————————1.16————————
PyramidInferModel=LLaMA-3.1-8B-Instruct, Prefill=60%, KV=60%, Execution strategy=Prefill-Aware2025.0239.32————————————————28.4119.3123.566.461.4136.87—————————
PyramidModel=Qwen2.5-32B, Eviction Metric=SnapKV Metric (π1), Compression ratio=80%2026.0238.68————————————————19.4440.2321.8460.2446.3450.46—————————
GemFilterModel=LLaMA-3.1-8B-Instruct, Prefill=51%, KV=10%, Execution strategy=Prefill-Aware2025.0238.61————————————————28.3940.3321.7864.3130.7546.13—————————
UniformModel=Qwen2.5-32B, Eviction Metric=KeyDiff Metric (π2), Compression ratio=80%2026.0238.33————————————————26.8544.322.3265.8229.3238.36—————————
H2OModel=LLaMA-3.1-8B-Instruct, Prefill=100%, KV=10%, Execution strategy=Decoding-Only2025.0238.15————————————————27.1219.7425.9666.6554.8834.58—————————
QUESTPer-Query Budget=1024, Backbone=Trado-8B-Instruct2026.0437.697.3717.35—39.88——————78.84———————————————————45
StreamingLLMBackbone=Qwen2.5-7B, Budget=10242026.0137.3822.7229.4231.4743.5738.1817.9924.3319.4722.466187.5343.798.53455.1758.43———————————————
Original LLMBackbone=Llama-2-7B, Latency=6.4, Speedup=4.0x2025.1237.2————————————————24.722.424.663.257.7——————————
StreamingLLMBackbone=Qwen2.5-7B, Budget=20482026.0137.0723.1836.9345.6445.340.119.7428.4920.5723.56874.4133.0681854.553.73———————————————
UniformModel=Mistral-7B-v0.3, Eviction Metric=KeyDiff Metric (π2), Compression ratio=80%2026.0236.83————————————————28.5330.8624.715946.8933.08—————————
QUESTPer-Query Budget=512, Backbone=Trado-8B-Instruct2026.0436.78.9916.57—34.58——————79.33———————————————————44.04
PyramidModel=Mistral-7B-v0.3, Eviction Metric=KeyDiff Metric (π2), Compression ratio=80%2026.0236.59————————————————29.4430.5324.559.3342.9534.06—————————
SpargeAttentionSparsity=τ=0.62025.1235.87——————————————————————0.98————————
QUESTPer-Query Budget=256, Backbone=Trado-8B-Instruct2026.0434.648.514.04—32.95——————75.23———————————————————42.46