Time to First Token (TTFT) on Generalization Models Tulu3, Qwen3, Lego-Link0
1,195.6TTFT (Standard, ms)LazyAttention
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| LazyAttentionModel=Lego-Link0 + Llama-3.1-8B2026.06 | 1,195.6 | 191.3 | 6.2 | |
| LazyAttentionModel=Tulu3-Block-FT 8B2026.06 | 1,196.8 | 191.7 | 6.2 | |
| LazyAttentionModel=Qwen3-8B2026.06 | 1,277.4 | 201.2 | 6.35 |