List your company

Speculative Decoding on Spec-Bench

195.6MT ScoreFR-Spec
-6.78445.75898.3150.842Nov 16, 2024Feb 10, 2025May 7, 2025Aug 1, 2025Oct 26, 2025Jan 20, 2026Apr 17, 2026

Evaluation results

MethodLinks
FR-SpecBase Model=Llama-3-8B, Temperature=0, Vocabulary Size (|Vhigh|)=32k, Implementation Framework=native C and CUDA2025.02195.6—188.14190.27243.36184.85—201.87—227.68183.192.27—
FR-SpecBase Model=Llama-3-8B, Temperature=0, Vocabulary Size (|Vhigh|)=16k, Implementation Framework=native C and CUDA2025.02194.02—182.26188.6233.69178.22—196.69—223.32176.72.21—
FR-SpecBase Model=Llama-3-8B, Temperature=0, Vocabulary Size (|Vhigh|)=64k, Implementation Framework=native C and CUDA2025.02192.85—183.86183.17231.99178.53—196.86—224.52183.112.21—
FR-SpecBase Model=Llama-3-8B, Temperature=0, Vocabulary Size (|Vhigh|)=8k, Implementation Framework=native C and CUDA2025.02185.78—170.97180.4218.57167.61—185.98—210.66167.852.09—
EAGLE-2Base Model=Llama-3-8B, Temperature=0, Implementation Framework=native C and CUDA2025.02176.79—167.12166.6209.88168.05—180.99—203.41175.112.03—
VanillaBase Model=Llama-3-8B, Temperature=0, Implementation Framework=native C and CUDA2025.0290.94—86.6391.0591.1683.43—89.11—90.4390.11—
PACERLLM Model=Llama-2 7B/70B2026.0218.5318.0616.1917.4221.4515.48——17.95————
SpecDec++LLM Model=Llama-2 7B/70B2026.0218.1817.6216.5717.4120.8115.8——17.79————
AssistLLM Model=Llama-2 7B/70B2026.0217.7717.9516.4217.120.4615.6——17.58————
AdaEDLLLM Model=Llama-2 7B/70B2026.0217.5517.8516.117.3420.2115.69——17.47————
VanillaLLM Model=Llama-2 7B/70B2026.029.049.318.639.329.298.32——8.99————
SAMDTemperature=0.72025.124.623.124.183.384.163.95——3.9————
Eagle-2Temperature=0.72025.124.543.193.813.414.383.8——3.86————
O-PTPTemperature=0.7, Base Model=Vicuna-7B2025.124.523.734.543.424.624.29——4.18————
HydraTemperature=0.72025.123.92.882.953.213.93.37——3.37————
EAGLE-3 + SpecVocabTarget Model=Qwen3 8B2026.023.82—4.494.926.285.26——5.015.3———
EAGLE-3 + FR-SpecTarget Model=Qwen3 8B2026.023.8—4.524.525.774.9——4.694.62———
EAGLE-3 + VocabTrimTarget Model=Qwen3 8B2026.023.69—4.284.726.045.13——4.845.17———
EagleTemperature=0.72025.123.662.743.152.813.573.21——3.19————
EAGLE-3 (Our Reproduction)Target Model=Qwen3 8B, Protocol=Reproduction2026.023.6—4.214.85.945.04——4.785.11———
EAGLE-3 + SpecVocab (Ours)Base Model=Qwen3 4B, Number of seeds=52026.023.59—4.635.096.055.45——5.015.26———
EAGLE-3 + FR-SpecBase Model=Qwen3 4B, Number of seeds=52026.023.52—4.594.665.614.79——4.624.57———
EAGLE-3 + SpecVocabBase Model=OLMo 2 1B2026.023.43—4.924.135.144.55——4.534.98———
EAGLE-3 + VocabTrimBase Model=Qwen3 4B, Number of seeds=52026.023.43—4.355.055.865.04——4.825.2———
EAGLE-3 + FR-SpecBase Model=OLMo 2 1B2026.023.41—4.774.044.874.32——4.334.56———
EAGLE-3 (Our Reproduction)Base Model=Qwen3 4B, Number of seeds=52026.023.37—4.314.945.794.94——4.735.02———
EAGLE-3Target Model=Qwen3 8B2026.023.34—3.864.465.284.52——4.334.52———
EAGLE-3Vocab=128K Full Vocab2026.043.34—3.623.413.894.15——3.784.29———
EvoSpecBase=Static 32k, Dynamic Expansion=Dyna Vocab 10242026.043.31—3.63.393.883.86——3.653.86———
EvoSpecBase=Static 32k, Dynamic Expansion=Dyna Vocab 5122026.043.3—3.583.383.883.86——3.643.85———
SAM-Decoding[E2]Model=Vicuna-13B2024.113.272.122.892.343.122.544.5157.882.72————
EAGLE-3 + VocabTrimBase Model=OLMo 2 1B2026.023.24—4.644.125.084.39——4.384.83———
EAGLE-3Base Model=OLMo 2 1B, Reproduction=true2026.023.22—4.5544.884.39——4.314.8———
EvoSpecBase=Static 32k, Dynamic Expansion=Dyna Vocab 2562026.043.22—3.543.373.813.82——3.63.81———
SAM-Decoding[E2]Model=Vicuna-7B, Hardware=A800 GPU, Draft Strategy=E22024.113.191.972.862.282.842.324.52129.362.58————
EAGLE-2Model=Vicuna-13B2024.113.12.152.582.383.192.334.4256.062.63————
FR-SpecVocab=Static 32k2026.043.1—3.513.293.793.81——3.543.73———
RecyclingTemperature=0.72025.122.742.482.682.593.052.63——2.7————
MedusaTemperature=0.72025.122.72.182.122.242.672.26——2.36————
EAGLE-2Model=Vicuna-7B, Hardware=A800 GPU2024.112.661.762.182.032.631.974.34110.562.21————
SAM-Decoding[T]Model=Vicuna-7B, Hardware=A800 GPU, Draft Strategy=T2024.112.621.822.922.092.62.213.02119.212.38————
SAM-Decoding[T]Model=Vicuna-13B2024.112.361.82.631.832.492.222.9147.272.22————
SpSTemperature=0.72025.122.121.392.161.81.922.26——1.94————
SAM-DecodingModel=Vicuna-13B2024.112.081.262.231.532.091.892.1939.241.84————
Token RecyclingModel=Vicuna-7B, Hardware=A800 GPU2024.112.081.761.971.852.351.762.8298.391.96————
Token RecyclingModel=Vicuna-13B2024.112.031.842.071.832.421.842.8142.742.01————
LogitSpecBackbone=Llama-3.1-8B-Instruct2025.071.891.671.941.682.011.772.11————1.88—
LogitSpecBackbone=Qwen-3-8B2025.071.711.741.641.651.891.681.95————1.75—
LookaheadTemperature=0.72025.121.691.241.541.561.921.42——1.56————
PLDTemperature=0.72025.121.671.12.731.371.81.71——1.73————
PLDModel=Vicuna-13B2024.111.611.12.361.111.691.81.6633.891.59————
NoneTemperature=0.72025.12111111——1————
Autoregressive DecodingBase Model=OLMo 2 1B2026.021—1111——11———
Autoregressive DecodingTarget Model=Qwen3 8B2026.021—1111——11———
Autoregressive DecodingBase Model=Qwen3 4B, Number of seeds=52026.021—1111——11———
VanillaBackbone=Llama-3.1-8B-Instruct2025.071111111————1—
VanillaBackbone=Qwen-3-8B2025.071111111————1—
CactusVerifier model=Qwen 3 14B, Drafter model=Qwen 3 0.6B, Temperature=0.6, delta=12026.04———————————2.093.29
EAGLEBackbone=Vicuna-7B, Note=Requires training & extra heads2025.07——————3.58————2.1—
LogitSpecBackbone=Vicuna 7B2025.07——————2.44————2.01—
LogitSpecBackbone=Vicuna 13B2025.07——————2.32————1.93—
LogitSpecBackbone=Vicuna 33B2025.07——————2.13————1.76—
LogitSpecBackbone=Vicuna-7B, Note=Training-free & plug-and-play2025.07——————2.44————2.01—
LookaheadBackbone=Vicuna 7B2025.07——————1.66————1.27—
LookaheadBackbone=Vicuna 13B2025.07——————1.63————1.22—
LookaheadBackbone=Vicuna 33B2025.07——————1.61————1.24—
MedusaBackbone=Vicuna-7B, Note=Requires training & extra heads2025.07——————2.31————1.71—
PLDBackbone=Vicuna 7B2025.07——————1.73————1.59—
PLDBackbone=Vicuna 13B2025.07——————1.68————1.48—
PLDBackbone=Vicuna 33B2025.07——————1.55————1.42—
RESTBackbone=Vicuna 7B2025.07——————1.82————1.48—
RESTBackbone=Vicuna 13B2025.07——————1.82————1.38—
RESTBackbone=Vicuna 33B2025.07——————1.8————1.48—
SpSVerifier model=Qwen 3 14B, Drafter model=Qwen 3 0.6B, Temperature=0.62026.04———————————2.013.2
SpSBackbone=Vicuna 7B2025.07——————2.28————1.49—
SpSBackbone=Vicuna 13B2025.07——————2.18————1.49—
SpSBackbone=Vicuna 33B2025.07——————2.01————1.61—