Long-context retrieval on RULER-NIAH 64k
98.8AccuracyBF16
Evaluation Results
| Method | Links | |
|---|---|---|
| BF16Backbone=GLM-4.7-FP8†2026.05 | 98.8 | |
| QuaRot-INT2Backbone=GLM-4.7-FP8†2026.05 | 98.8 | |
| OSCARBackbone=GLM-4.7-FP8†2026.05 | 98.8 | |
| BF16Backbone=Qwen3-4B-Thinking-25072026.05 | 85.3 | |
| BF16Backbone=Qwen3-8B2026.05 | 79.2 | |
| OSCARBackbone=Qwen3-4B-Thinking-25072026.05 | 61.9 | |
| OSCARBackbone=Qwen3-8B2026.05 | 61.9 | |
| QuaRot-INT2Backbone=Qwen3-4B-Thinking-25072026.05 | 15.6 | |
| QuaRot-INT2Backbone=Qwen3-8B2026.05 | 0 |