Long-Context Generation on Qwen3 Context length (50K)
6.02Throughput Speedup (α)SpecPV
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SpecPVModel Size=14B, Partial KV Cache Budget=4K2025.12 | 6.02 | 3.69 | |
| SpecPVModel Size=14B, Partial KV Cache Budget=2K2025.12 | 5.99 | 3.6 | |
| SpecPVModel Size=14B, Partial KV Cache Budget=8K2025.12 | 5.45 | 3.57 | |
| SpecPVModel Size=4B, Partial KV Cache Budget=4K2025.12 | 5.44 | 3.28 | |
| SpecPVModel Size=4B, Partial KV Cache Budget=2K2025.12 | 5.32 | 3.15 | |
| SpecPVModel Size=8B, Partial KV Cache Budget=2K2025.12 | 5.24 | 3.32 | |
| SpecPVModel Size=8B, Partial KV Cache Budget=4K2025.12 | 5.21 | 3.29 | |
| SpecPVModel Size=4B, Partial KV Cache Budget=8K2025.12 | 5.04 | 3.11 | |
| SpecPVModel Size=8B, Partial KV Cache Budget=8K2025.12 | 4.97 | 3.27 | |
| EAGLE3-YARNModel Size=14B2025.12 | 3.18 | 3.78 | |
| EAGLE3-YARNModel Size=8B2025.12 | 2.78 | 3.29 | |
| EAGLE3-YARNModel Size=4B2025.12 | 2.7 | 3.13 |