List your company

Multi-step LLM Interaction on Qwen3-4B Inference Performance (test)

100High-Precision RatioBF16
-4235077Feb 2, 2026

Evaluation results

MethodLinks
BF16Precision=BF162026.0210093.315.7
Random@80%Sampling Ratio=80%2026.028090.315.88
Random@60%Sampling Ratio=60%2026.026088.818.67
Random@40%Sampling Ratio=40%2026.024081.18.75
Random@20%Sampling Ratio=20%2026.022078.12.5
Router(Ours)Routing=Dynamic Mix Precision2026.028.681.343.02
GPTQ-3BitQuantization=3-bit2026.02077.6—