Multi-turn Dialogue Processing on BCP
17.1Throughput (k)Full-cache
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Full-cacheBackbone=Qwen2.5-14B, Capacity (C)=16k2026.06 | 17.1 | 203.8 | 25.2 | 27.1 | |
| H2OBackbone=Qwen2.5-14B, Capacity (C)=16k2026.06 | 15.2 | 521.3 | 80.9 | 58.7 | |
| IntentKV-Phase-2Backbone=Qwen2.5-14B, Capacity (C)=16k2026.06 | 15.1 | 105 | 28.8 | 24.1 | |
| SnapKVBackbone=Qwen2.5-14B, Capacity (C)=16k2026.06 | 14.9 | 447.5 | 70.7 | 52.6 | |
| StreamingLLMBackbone=Qwen2.5-14B, Capacity (C)=16k2026.06 | 14.5 | 292.4 | 64.5 | 51.8 | |
| SnapKVBackbone=Qwen2.5-14B, Capacity (C)=8k2026.06 | 12.7 | 574.2 | 101 | 56.5 | |
| H2OBackbone=Qwen2.5-14B, Capacity (C)=8k2026.06 | 12.6 | 373.1 | 93.6 | 52.4 | |
| IntentKV-Phase-2Backbone=Qwen2.5-14B, Capacity (C)=8k2026.06 | 11.8 | 104 | 26.7 | 21 | |
| Full-cacheBackbone=Qwen3-8B, Capacity (C)=16k2026.06 | 10.6 | 131.4 | 32.2 | 45.1 | |
| StreamingLLMBackbone=Qwen2.5-14B, Capacity (C)=8k2026.06 | 10.5 | 275.3 | 68.3 | 44.5 | |
| H2OBackbone=Qwen3-8B, Capacity (C)=16k2026.06 | 10.2 | 244.7 | 65.5 | 63.6 | |
| StreamingLLMBackbone=Qwen3-8B, Capacity (C)=16k2026.06 | 9.9 | 165.3 | 50 | 51.5 | |
| SnapKVBackbone=Qwen3-8B, Capacity (C)=16k2026.06 | 9.9 | 228.2 | 60.7 | 59.1 | |
| IntentKV-Phase-2Backbone=Qwen3-8B, Capacity (C)=16k2026.06 | 9.9 | 139.2 | 32 | 43.2 | |
| H2OBackbone=Qwen3-8B, Capacity (C)=8k2026.06 | 9.1 | 281.4 | 77.2 | 59.7 | |
| SnapKVBackbone=Qwen3-8B, Capacity (C)=8k2026.06 | 9 | 279.2 | 77.1 | 59.4 | |
| StreamingLLMBackbone=Qwen3-8B, Capacity (C)=8k2026.06 | 8.4 | 180.6 | 54.3 | 45.2 | |
| IntentKV-Phase-2Backbone=Qwen3-8B, Capacity (C)=8k2026.06 | 8.1 | 140.4 | 31.6 | 39.8 |