Needle-in-a-Haystack retrieval on NIAH-Single 128K context 2
98.6AccuracyQwen3+YaRN
Evaluation Results
| Method | Links | |
|---|---|---|
| Qwen3+YaRNBackbone=8B, LS. Tokens=-2026.06 | 98.6 | |
| FlashMorphBackbone=1.7B, LS. Tokens=20M2026.06 | 98.2 | |
| HALOBackbone=1.7B, LS. Tokens=234M2026.06 | 95 | |
| HALOBackbone=8B, LS. Tokens=234M2026.06 | 89.2 | |
| FlashMorphBackbone=8B, LS. Tokens=20M2026.06 | 85.6 | |
| Qwen3+YaRNBackbone=30B-A3B, LS. Tokens=-2026.06 | 79.6 | |
| PostNAS*Backbone=1.7B, LS. Tokens=50B2026.06 | 78 | |
| HALOBackbone=0.6B, LS. Tokens=234M2026.06 | 62.4 | |
| HALOBackbone=30B-A3B, LS. Tokens=234M2026.06 | 61.2 | |
| KL-LSBackbone=0.6B, LS. Tokens=20B2026.06 | 50.2 | |
| KL-LSBackbone=1.7B, LS. Tokens=20B2026.06 | 47.4 | |
| FlashMorphBackbone=0.6B, LS. Tokens=20M2026.06 | 45.8 | |
| UniformBackbone=0.6B, LS. Tokens=N/A2026.06 | 41.2 | |
| UniformBackbone=30B-A3B, LS. Tokens=N/A2026.06 | 35.4 | |
| UniformBackbone=8B, LS. Tokens=N/A2026.06 | 31.6 | |
| UniformBackbone=1.7B, LS. Tokens=N/A2026.06 | 28.4 | |
| FlashMorphBackbone=30B-A3B, LS. Tokens=20M2026.06 | 21.4 | |
| Qwen3+YaRNBackbone=1.7B, LS. Tokens=-2026.06 | 1.8 | |
| Qwen3Backbone=0.6B, LS. Tokens=-2026.06 | 0 | |
| Qwen3+YaRNBackbone=0.6B, LS. Tokens=-2026.06 | 0 | |
| Qwen3Backbone=1.7B, LS. Tokens=-2026.06 | 0 | |
| Qwen3Backbone=8B, LS. Tokens=-2026.06 | 0 | |
| Qwen3Backbone=30B-A3B, LS. Tokens=-2026.06 | 0 |