List your company

Inference Acceleration on Spec-Bench

5.31SpeedupCATS w/ EAGLE
1.03562.14533.2554.3647Nov 16, 2024Feb 14, 2025May 15, 2025Aug 13, 2025Nov 11, 2025Feb 9, 2026May 11, 2026

Evaluation results

MethodLinks
CATS w/ EAGLEModel=Llama2-7b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=tree base2026.055.31——————————5.3143
CATS w/ EAGLEModel=Llama2-13b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=tree base2026.055.01——————————5.0813
CATSModel=Llama2-7b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=chain base2026.054.88——————————4.8831
CATS w/ EAGLEModel=Llama2-7b, Decoding Strategy=Greedy, Decoding Structure=tree base2026.054.83——————————4.8318
EAGLE-self, treeModel=Llama2-13b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=tree base2026.054.67——————————4.7176
CATSModel=Llama2-13b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=chain base2026.054.61——————————4.6461
CATS w/ EAGLEModel=Llama2-13b, Decoding Strategy=Greedy, Decoding Structure=tree base2026.054.51——————————4.5707
CATSModel=Llama2-7b, Decoding Strategy=Greedy, Decoding Structure=chain base2026.054.33——————————4.3335
EAGLE-self, treeModel=Llama2-7b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=tree base2026.054.29——————————4.9713
KangarooModel=Llama2-13b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=chain base2026.054.14——————————4.1486
EAGLE-selfModel=Llama2-13b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=chain base2026.054.12——————————4.1452
EAGLE-self, treeModel=Llama2-13b, Decoding Strategy=Greedy, Decoding Structure=tree base2026.054.04——————————4.0829
CATSModel=Llama2-13b, Decoding Strategy=Greedy, Decoding Structure=chain base2026.054.02——————————4.057
CATS w/ EAGLEModel=Vicuna-13b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=tree base2026.053.93——————————3.9851
KangarooModel=Llama2-7b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=chain base2026.053.87——————————4.4849
EAGLE-selfModel=Llama2-7b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=chain base2026.053.87——————————4.491
EAGLE-self, treeModel=Llama2-7b, Decoding Strategy=Greedy, Decoding Structure=tree base2026.053.75——————————4.3529
KangarooModel=Llama2-13b, Decoding Strategy=Greedy, Decoding Structure=chain base2026.053.53——————————3.5409
EAGLE-selfModel=Llama2-13b, Decoding Strategy=Greedy, Decoding Structure=chain base2026.053.53——————————3.5478
CATS w/ EAGLEModel=Vicuna-13b, Decoding Strategy=Greedy, Decoding Structure=tree base2026.053.42——————————3.4708
KangarooModel=Llama2-7b, Decoding Strategy=Greedy, Decoding Structure=chain base2026.053.29——————————3.8176
EAGLE-selfModel=Llama2-7b, Decoding Strategy=Greedy, Decoding Structure=chain base2026.053.28——————————3.8068
EAGLE-self, treeModel=Vicuna-13b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=tree base2026.053.16——————————3.1934
DECODevice=Jetson AGX 64GB, Kernel=Ours (acceleration kernel)2026.05344.1544.0842.7845.2945.9943.64———44.32—
CATSModel=Vicuna-13b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=chain base2026.052.95——————————2.9743
EAGLE-self, treeModel=Vicuna-13b, Decoding Strategy=Greedy, Decoding Structure=tree base2026.052.77——————————2.8035
CATSModel=Vicuna-13b, Decoding Strategy=Greedy, Decoding Structure=chain base2026.052.64——————————2.6588
SAM-Decoding[E2]Backbone=Vicuna-7B-v1.3, Auxiliary Method=EAGLE-2, Data Type=float16, Batch size=12024.112.58——————4.6297.56———
DECODevice=RTX 4090 24GB, Kernel=Ours (acceleration kernel)2026.052.58224.28222.65218.24228.57231.88222.19———224.63—
EAGLE-selfModel=Vicuna-13b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=chain base2026.052.5——————————2.5144
KangarooModel=Vicuna-13b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=chain base2026.052.48——————————2.4889
EAGLE-2Backbone=Vicuna-7B-v1.3, Data Type=float16, Batch size=12024.112.38——————4.3690.14———
SAM-Decoding[T]Backbone=Vicuna-7B-v1.3, Auxiliary Method=Token Recycling, Data Type=float16, Batch size=12024.112.27——————3.0385.73———
MedusaModel=Vicuna-13b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=chain base2026.052.2——————————2.5852
KangarooModel=Vicuna-13b, Decoding Strategy=Greedy, Decoding Structure=chain base2026.052.19——————————2.1937
EAGLE-selfModel=Vicuna-13b, Decoding Strategy=Greedy, Decoding Structure=chain base2026.052.18——————————2.1952
SAM-Decoding[E2]Backbone=Llama3-8B, Combined Method=EAGLE-22024.112.082.361.961.981.792.322.113.9272.47———
SAM-Decoding[T]Backbone=Llama3-8B, Combined Method=Token Recycling2024.112.052.091.932.041.822.322.122.6371.73———
MedusaModel=Vicuna-13b, Decoding Strategy=Greedy, Decoding Structure=chain base2026.052.03——————————2.392
EAGLE-2Backbone=Llama3-8B2024.111.982.081.951.851.82.311.873.968.69———
Token RecyclingBackbone=Llama3-8B2024.111.911.921.881.851.752.241.822.7666.42———
SAM-DecodingBackbone=Vicuna-7B-v1.3, Data Type=float16, Batch size=12024.111.84——————2.369.37———
Token RecyclingBackbone=Vicuna-7B-v1.3, Data Type=float16, Batch size=12024.111.84——————2.8369.65———
MedusaModel=Llama2-7b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=chain base2026.051.75——————————1.7482
MedusaModel=Llama2-7b, Decoding Strategy=Greedy, Decoding Structure=chain base2026.051.67——————————1.666
MedusaModel=Llama2-13b, Decoding Strategy=Relaxed (temperature = 0.7), Decoding Structure=chain base2026.051.64——————————1.9322
PLDBackbone=Vicuna-7B-v1.3, Data Type=float16, Batch size=12024.111.56——————1.7559.02———
MedusaModel=Llama2-13b, Decoding Strategy=Greedy, Decoding Structure=chain base2026.051.55——————————1.825
SAM-DecodingBackbone=Llama3-8B2024.111.511.591.351.51.351.541.751.7252.35———
PIABackbone=Vicuna-7B-v1.3, Data Type=float16, Batch size=12024.111.47——————2.0855.45———
REST*Backbone=Vicuna-7B-v1.3, Data Type=float16, Batch size=12024.111.38——————1.6351.34———
PLDBackbone=Llama3-8B2024.111.281.31.121.411.031.31.531.3944.26———
Lookahead*Backbone=Vicuna-7B-v1.3, Data Type=float16, Batch size=12024.111.2——————1.6344.37———
ARParameters=7B2024.08—111111154.3———
ARParameters=13B2024.08—111111139.41———
ARParameters=33B2024.08—111111118.44———
Baseline ARDevice=RTX 4090 24GB, Kernel=Standard autoregressive decoding2026.05—86.6387.8186.287.8887.8386.24———87.1—
Baseline ARDevice=Jetson AGX 64GB, Kernel=Standard autoregressive decoding2026.05—14.6814.8914.6114.9114.8914.62———14.77—
EAGLE-2Model=Vicuna-33B2024.11—3.292.312.732.513.652.464.0625.862.83——
LadeParameters=7B2024.08—1.421.121.211.211.521.131.6469.03———
LadeParameters=13B2024.08—1.291.061.161.121.481.091.6347.5———
LadeParameters=33B2024.08—1.321.091.21.171.551.141.6123.03———
MedusaParameters=7B2024.08—1.91.571.481.581.871.452.3189.41———
MedusaParameters=13B2024.08—1.941.661.571.621.981.532.3967.92———
MedusaParameters=33B2024.08—1.981.751.631.682.091.612.3233.11———
PLDModel=Vicuna-33B2024.11—1.51.072.061.091.591.511.6513.331.46——
PLDParameters=7B2024.08—1.530.982.361.11.51.741.7583.3———
PLDParameters=13B2024.08—1.451.012.11.021.551.651.6757.01———
PLDParameters=33B2024.08—1.431.061.941.081.551.411.5525.89———
RESTParameters=7B2024.08—1.371.051.121.421.061.31.8466.29———
RESTParameters=13B2024.08—1.511.141.311.51.171.51.8253.34———
RESTParameters=33B2024.08—1.631.271.421.611.291.571.8126.99———
SAM-DecodingModel=Vicuna-33B2024.11—1.911.251.981.481.831.661.9715.351.68——
SAM-DecodingModel=Vicuna-33B, variant=T2024.11—2.311.792.531.92.482.062.6819.872.18——
SAM-DecodingModel=Vicuna-33B, variant=E22024.11—3.42.252.932.433.452.544.0825.912.84——
Token RecyclingModel=Vicuna-33B2024.11—2.11.842.191.882.421.922.718.82.06——
TRParameters=7B2024.08—2.171.91.941.952.41.782.7110.06———
TRParameters=13B2024.08—1.981.771.891.752.211.732.7274.57———
TRParameters=33B2024.08—1.951.751.921.772.241.782.6335.16———