Long-Context Inference on LongBench >128K tokens v2 (long)
35.2AccuracySnapKV (Best)
Evaluation Results
| Method | Links | |
|---|---|---|
| SnapKV (Best)KV Cache Budget=4096 tokens2026.02 | 35.2 | |
| CHESS (Aggressive)KV Cache Budget=1%2026.02 | 33.3 | |
| KeyDiff (Best)KV Cache Budget=8192 tokens2026.02 | 32.4 | |
| H2O (Best)KV Cache Budget=20%2026.02 | 31.5 | |
| Quest (Best)KV Cache Budget=2048 tokens2026.02 | 31.5 | |
| CHESS (Moderate)KV Cache Budget=40%2026.02 | 31.5 | |
| CHESS (Conservative)KV Cache Budget=73%2026.02 | 29.6 | |
| FullKV InferenceKV Cache Budget=100%2026.02 | 28.7 |