| Dataset Name | SOTA Method | Metric | Trend | ||
|---|---|---|---|---|---|
| Qwen3-8B Single-layer forward+backward setup | DualKV | Time (ms)36.6 | 57 | 2mo ago | |
| Llama 8B base variant on H200 3.1 | FORGE | Per-step Latency (ms)99.1 | 49 | 1mo ago | |
| ImageNet-1K 1.0 (test) | Isomorphic Pruning | GPU Latency (ms)17.8 | 12 | 5mo ago | |
| DeepSeek-R1-Distill-Llama-8B 16K Generation Length | SnapKV | Max Batch Size8 | 10 | 1mo ago | |
| DeepSeek-R1-Distill-Llama-8B 8K Generation Length | SnapKV | Max Batch Size8 | 10 | 1mo ago | |
| 160s audio context | MFCCT | Latency Gain3.65 | 4 | 1mo ago |
