Long-context recall on NIAH Single-3
100Recall @ 32K ContextQwen3
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Qwen3Param=1.7B, Token=-, Evaluation Protocol=Evaluated with YaRN, Architecture=teacher, no RNNs2026.01 | 100 | 98.4 | 14.8 | 19 | |
| KL-LS (GDN)Param=3B, Token=25B2026.01 | 99 | 51 | 24.8 | 11 | |
| HypeNet + HALOParam=2B, Token=2.3B2026.01 | 87.2 | 72.6 | 44.8 | 48.8 | |
| Jet-NemotronParam=2B, Token=400B2026.01 | 84 | 15.4 | 0 | 0 |