Long-context recall on NIAH Single-1
100Recall @ 32KQwen3
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Qwen3Param=1.7B, Token=-, Evaluation Protocol=Evaluated with YaRN, Architecture=teacher, no RNNs2026.01 | 100 | 100 | 96.4 | 17 | |
| Jet-NemotronParam=2B, Token=400B2026.01 | 99.8 | 56 | 0 | 0 | |
| KL-LS (GDN)Param=3B, Token=25B2026.01 | 99.8 | 99.4 | 68.4 | 14.8 | |
| HypeNet + HALOParam=2B, Token=2.3B2026.01 | 99.8 | 99.6 | 99.8 | 99.8 |