Long-Context Generation on Qwen3 Context length 20K
3.73Throughput Speedup (α)SpecPV
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SpecPVModel Size=14B, Partial KV Cache Budget=2K2025.12 | 3.73 | 3.08 | |
| SpecPVModel Size=4B, Partial KV Cache Budget=2K2025.12 | 3.69 | 3.31 | |
| SpecPVModel Size=14B, Partial KV Cache Budget=4K2025.12 | 3.68 | 3.14 | |
| SpecPVModel Size=8B, Partial KV Cache Budget=2K2025.12 | 3.58 | 3.05 | |
| SpecPVModel Size=4B, Partial KV Cache Budget=8K2025.12 | 3.5 | 3.3 | |
| SpecPVModel Size=8B, Partial KV Cache Budget=4K2025.12 | 3.47 | 3.06 | |
| SpecPVModel Size=4B, Partial KV Cache Budget=4K2025.12 | 3.4 | 3.11 | |
| SpecPVModel Size=14B, Partial KV Cache Budget=8K2025.12 | 3.35 | 3.11 | |
| SpecPVModel Size=8B, Partial KV Cache Budget=8K2025.12 | 3.33 | 3.19 | |
| EAGLE3-YARNModel Size=14B2025.12 | 2.74 | 3.3 | |
| EAGLE3-YARNModel Size=8B2025.12 | 2.65 | 3.25 | |
| EAGLE3-YARNModel Size=4B2025.12 | 2.59 | 3.28 |