Needle-in-a-Haystack retrieval on NIAH-Single-3 64K context
100AccuracyQwen3
Evaluation Results
| Method | Links | |
|---|---|---|
| Qwen3Backbone=30B-A3B, LS. Tokens=-2026.06 | 100 | |
| Qwen3Backbone=8B, LS. Tokens=-2026.06 | 99.8 | |
| FlashMorphBackbone=8B, LS. Tokens=20M2026.06 | 98.2 | |
| Qwen3+YaRNBackbone=8B, LS. Tokens=-2026.06 | 98 | |
| Qwen3Backbone=1.7B, LS. Tokens=-2026.06 | 96.4 | |
| FlashMorphBackbone=1.7B, LS. Tokens=20M2026.06 | 95.4 | |
| HALOBackbone=1.7B, LS. Tokens=234M2026.06 | 90.8 | |
| HALOBackbone=8B, LS. Tokens=234M2026.06 | 85.2 | |
| Qwen3Backbone=0.6B, LS. Tokens=-2026.06 | 82.8 | |
| FlashMorphBackbone=0.6B, LS. Tokens=20M2026.06 | 73.6 | |
| HALOBackbone=0.6B, LS. Tokens=234M2026.06 | 67.4 | |
| UniformBackbone=1.7B, LS. Tokens=N/A2026.06 | 59.4 | |
| UniformBackbone=8B, LS. Tokens=N/A2026.06 | 54.2 | |
| PostNAS*Backbone=1.7B, LS. Tokens=50B2026.06 | 51.2 | |
| FlashMorphBackbone=30B-A3B, LS. Tokens=20M2026.06 | 38.2 | |
| UniformBackbone=0.6B, LS. Tokens=N/A2026.06 | 36 | |
| UniformBackbone=30B-A3B, LS. Tokens=N/A2026.06 | 36 | |
| HALOBackbone=30B-A3B, LS. Tokens=234M2026.06 | 17.8 | |
| KL-LSBackbone=0.6B, LS. Tokens=20B2026.06 | 17.2 | |
| Qwen3+YaRNBackbone=30B-A3B, LS. Tokens=-2026.06 | 17.2 | |
| KL-LSBackbone=1.7B, LS. Tokens=20B2026.06 | 4 | |
| Qwen3+YaRNBackbone=1.7B, LS. Tokens=-2026.06 | 1 | |
| Qwen3+YaRNBackbone=0.6B, LS. Tokens=-2026.06 | 0 |