Long-Context Generation on Qwen3 Context length 10K
2.76Throughput Speedup (α)SpecPV
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| SpecPVModel Size=4B, Partial KV Cache Budget=2K2025.12 | 2.76 | 3.39 | |
| SpecPVModel Size=14B, Partial KV Cache Budget=4K2025.12 | 2.75 | 2.89 | |
| SpecPVModel Size=14B, Partial KV Cache Budget=2K2025.12 | 2.74 | 2.79 | |
| SpecPVModel Size=4B, Partial KV Cache Budget=4K2025.12 | 2.55 | 3.37 | |
| SpecPVModel Size=8B, Partial KV Cache Budget=2K2025.12 | 2.53 | 2.9 | |
| SpecPVModel Size=4B, Partial KV Cache Budget=8K2025.12 | 2.46 | 3.35 | |
| EAGLE3-YARNModel Size=14B2025.12 | 2.46 | 3.03 | |
| SpecPVModel Size=14B, Partial KV Cache Budget=8K2025.12 | 2.46 | 2.91 | |
| EAGLE3-YARNModel Size=8B2025.12 | 2.44 | 3.23 | |
| SpecPVModel Size=8B, Partial KV Cache Budget=4K2025.12 | 2.44 | 2.98 | |
| EAGLE3-YARNModel Size=4B2025.12 | 2.43 | 3.27 | |
| SpecPVModel Size=8B, Partial KV Cache Budget=8K2025.12 | 2.39 | 3.02 |