Long-Context Generation on Qwen3 Context length 30K
4.37Throughput Speedup (α)SpecPV
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SpecPVModel Size=14B, Partial KV Cache Budget=2K2025.12 | 4.37 | 3.1 | |
| SpecPVModel Size=4B, Partial KV Cache Budget=2K2025.12 | 4.31 | 3.25 | |
| SpecPVModel Size=4B, Partial KV Cache Budget=4K2025.12 | 4.22 | 3.17 | |
| SpecPVModel Size=8B, Partial KV Cache Budget=2K2025.12 | 4.14 | 3.01 | |
| SpecPVModel Size=8B, Partial KV Cache Budget=4K2025.12 | 4.13 | 3.02 | |
| SpecPVModel Size=14B, Partial KV Cache Budget=4K2025.12 | 4.13 | 3.02 | |
| SpecPVModel Size=4B, Partial KV Cache Budget=8K2025.12 | 4.02 | 3.23 | |
| SpecPVModel Size=8B, Partial KV Cache Budget=8K2025.12 | 4 | 3.19 | |
| SpecPVModel Size=14B, Partial KV Cache Budget=8K2025.12 | 3.83 | 2.98 | |
| EAGLE3-YARNModel Size=14B2025.12 | 2.83 | 3.34 | |
| EAGLE3-YARNModel Size=8B2025.12 | 2.79 | 3.35 | |
| EAGLE3-YARNModel Size=4B2025.12 | 2.61 | 3.21 |