Long-context retrieval on RULER-NIAH 128k
97.2AccuracyBF16
Evaluation Results
| Method | Links | |
|---|---|---|
| BF16Backbone=GLM-4.7-FP8†2026.05 | 97.2 | |
| OSCARBackbone=GLM-4.7-FP8†2026.05 | 97.2 | |
| QuaRot-INT2Backbone=GLM-4.7-FP8†2026.05 | 96.3 | |
| BF16Backbone=Qwen3-4B-Thinking-25072026.05 | 81 | |
| BF16Backbone=Qwen3-8B2026.05 | 78.2 | |
| OSCARBackbone=Qwen3-8B2026.05 | 45 | |
| OSCARBackbone=Qwen3-4B-Thinking-25072026.05 | 39.5 | |
| QuaRot-INT2Backbone=Qwen3-4B-Thinking-25072026.05 | 0 | |
| QuaRot-INT2Backbone=Qwen3-8B2026.05 | 0 |