List your company

Multi-step LLM Interaction on Qwen3-1.7B Inference Performance (test)

100High-Precision RatioBF16
-4235077Feb 2, 2026

Evaluation results

MethodLinks
BF16Precision=BF162026.0210069.414.9
Random@80%Sampling Ratio=80%2026.028065.714
Random@60%Sampling Ratio=60%2026.026061.912.33
Random@40%Sampling Ratio=40%2026.024060.515
Router(Ours)Routing=Dynamic Mix Precision2026.0220.55921.95
Random@20%Sampling Ratio=20%2026.022056.28.5
GPTQ-Int4Quantization=4-bit2026.02054.5—