List your company

Multi-step LLM Interaction on Qwen3-8B (test)

100High-Precision RatioBF16
-4235077Feb 2, 2026

Evaluation results

MethodLinks
BF16Precision=BF162026.0210089.66.8
Random@80%Sampling Ratio=80%2026.028087.76.13
Random@60%Sampling Ratio=60%2026.026084.833.38
Random@40%Sampling Ratio=40%2026.024086.28.5
Router(Ours)Routing=Dynamic Mix Precision2026.0226.788.819.85
Random@20%Sampling Ratio=20%2026.022083.22
GPTQ-3BitQuantization=3-bit2026.02082.8—