Multi-key Needle-in-a-Haystack Recall on 32k Context Length
100Recallfp16 baseline
Evaluation Results
| Method | Links | |
|---|---|---|
| fp16 baselinebits=–, codec=fp16 baseline, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 100 | |
| TurboQuant-MSEbits=4, codec=TurboQuant-MSE, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 100 | |
| TurboQuant-QJLbits=4, codec=TurboQuant-QJL, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 100 | |
| PolarQuantbits=4, codec=PolarQuant, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 100 | |
| OCTOPUSbits=4, codec=OCTOPUS, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 100 | |
| OCTOPUS-QJLbits=4, codec=OCTOPUS-QJL, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 100 | |
| TurboQuant-MSEbits=3, codec=TurboQuant-MSE, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 100 | |
| OCTOPUSbits=3, codec=OCTOPUS, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 100 | |
| OCTOPUS-QJLbits=3, codec=OCTOPUS-QJL, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 100 | |
| PolarQuantbits=3, codec=PolarQuant, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 90 | |
| OCTOPUS-QJLbits=2, codec=OCTOPUS-QJL, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 85 | |
| OCTOPUSbits=2, codec=OCTOPUS, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 75 | |
| TurboQuant-QJLbits=3, codec=TurboQuant-QJL, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 60 | |
| TurboQuant-MSEbits=2, codec=TurboQuant-MSE, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 50 | |
| PolarQuantbits=2, codec=PolarQuant, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 5 | |
| TurboQuant-QJLbits=2, codec=TurboQuant-QJL, Backbone=Qwen2.5-7B-Instruct-1M2026.05 | 0 |