Long-context Retrieval on RULER 8k context NIAH Single 1
100AccuracyBF16
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| BF16Backbone=Qwen3-4B-Thinking-25072026.05 | 100 | — | — | — | |
| OSCARBackbone=Qwen3-4B-Thinking-25072026.05 | 100 | — | — | — | |
| BF16Backbone=GLM-4.7-FP8†2026.05 | 100 | — | — | — | |
| OSCARBackbone=GLM-4.7-FP8†2026.05 | 100 | — | — | — | |
| QuaRot-INT2Backbone=GLM-4.7-FP8†2026.05 | 99.7 | — | — | — | |
| BF16Backbone=Qwen3-8B2026.05 | 99.6 | — | — | — | |
| OSCARBackbone=Qwen3-8B2026.05 | 97.8 | — | — | — | |
| QuaRot-INT2Backbone=Qwen3-8B2026.05 | 80.6 | — | — | — | |
| QuaRot-INT2Backbone=Qwen3-4B-Thinking-25072026.05 | 0.8 | — | — | — | |
| Self-Pruned KVAttention Mechanism=Self-Pruned KV, tau=0.5, Model Size=8.1B, Training Context=32k2026.05 | — | 1 | 0 | 5.6 | |
| VanillaAttention Mechanism=Full Attention, Model Size=8.1B, Training Context=32k2026.05 | — | 1 | — | — |