Long-context understanding on LongBench V1 (Full Metric Set)
45.77QasperSlow
Evaluation Results
| Method | Links | ||||||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SlowBackbone=Qwen3-235B-A22B, Decoding method=full-KV decoding2026.03 | 45.77 | 51.21 | 67.08 | 67.13 | 64.14 | 42.86 | 25.38 | 31.68 | 22.81 | 23.46 | 77.5 | 91.86 | 41.09 | 51 | 100 | 61.32 | 62.56 | 54.52 | |
| SFIBackbone=Qwen3-235B-A22B, Decoding method=Slow-Fast Inference2026.03 | 44.67 | 50.92 | 66.37 | 67.65 | 65.31 | 43.44 | 24.58 | 31.28 | 22.72 | 23.33 | 77.5 | 92.1 | 41.3 | 52 | 100 | 61.1 | 63.18 | 54.56 | |
| SFIBackbone=Qwen3-4B, Decoding method=Slow-Fast Inference2026.03 | 44.2 | 49.31 | 63.81 | 59 | 44.5 | 25.76 | 23.77 | 29.72 | 22.21 | 24.04 | 75 | 85.62 | 39.99 | 37.75 | 100 | 4.32 | 5.28 | 43.19 | |
| SFIBackbone=Qwen3-30B-A3B, Decoding method=Slow-Fast Inference2026.03 | 42.14 | 53.22 | 66 | 63.37 | 55.98 | 31.95 | 23.44 | 30.18 | 21.92 | 23.46 | 78.5 | 91.06 | 39.72 | 47.25 | 100 | 25.13 | 24.92 | 48.13 | |
| SlowBackbone=Qwen3-4B, Decoding method=full-KV decoding2026.03 | 40.6 | 46.56 | 61.21 | 55.34 | 42.02 | 24.79 | 21.85 | 27.87 | 22.11 | 24.06 | 73 | 85.29 | 39.12 | 30.25 | 100 | 4.48 | 5.17 | 41.4 | |
| SlowBackbone=Qwen3-30B-A3B, Decoding method=full-KV decoding2026.03 | 38.96 | 50.32 | 63.42 | 61.68 | 54.68 | 32.22 | 21.4 | 29.4 | 21.66 | 23.52 | 77.5 | 91.56 | 39.12 | 42.5 | 100 | 24.82 | 24.76 | 46.91 |