List your company

Long-context Retrieval on NIAH (Needle In A Haystack) Multi-task Qwen2.5-7B

96Accuracy (Single)fp16
-3.8422.084873.92Jun 23, 2026

Evaluation results

MethodLinks
fp16Backbone=Qwen2.5-7B-Instruct, Rate=fp162026.069692.983.831.566.33267.1
Block-GTQBackbone=Qwen2.5-7B-Instruct, Rate=K3V3, Calibration=2048-token WikiText-22026.069691.478.830.66231.865.1
Block-GTQBackbone=Qwen2.5-7B-Instruct, Rate=K3V2, Calibration=2048-token WikiText-22026.0692.990.975.337.461.330.864.8
Block-GTQBackbone=Qwen2.5-7B-Instruct, Rate=K2V2, Calibration=2048-token WikiText-22026.0685.977.873.240.154.729.160.1
KIVI-ScaleOnlyBackbone=Qwen2.5-7B-Instruct, Rate=K3V32026.0659.144.437.920.732.916.335.2
KIVI-ScaleOnlyBackbone=Qwen2.5-7B-Instruct, Rate=K3V22026.0659.142.424.823.224.119.532.2
KIVI-ScaleOnlyBackbone=Qwen2.5-7B-Instruct, Rate=K2V22026.0620.714.110.612.16.67.912
TQ-MSEBackbone=Qwen2.5-7B-Instruct, Rate=K3V32026.060000000
TQ-MSEBackbone=Qwen2.5-7B-Instruct, Rate=K3V22026.060000000
TQ-MSEBackbone=Qwen2.5-7B-Instruct, Rate=K2V22026.060000000