List your company

Speculative Decoding on LWM 16K context

1.94SpeedupBudgetDraft
0.73361.04681.361.6732May 29, 2026

Evaluation results

MethodLinks
BudgetDraftDrafter=68M, KV budget (B)=256, Draft length (γ)=52026.051.94
BudgetDraftDrafter=68M, KV budget (B)=512, Draft length (γ)=52026.051.89
BudgetDraftDrafter=68M, KV budget (B)=1024, Draft length (γ)=52026.051.77
BudgetDraftDrafter=68M, KV budget (B)=2048, Draft length (γ)=52026.051.52
EAGLE-3Drafter=draft head, Draft length (γ)=52026.051.36
TriForceDrafter=68M+7B, Draft length (γ)=52026.051.19
ARDrafter=–, Draft length (γ)=52026.051
SD (sparse/full)Drafter=68M, Draft length (γ)=52026.050.78