Analysis of Mixture-of-Experts routing in Qwen 3.5 and 3.6 35B models using MT-Bench prompts, examining which experts are activated per token and the router's confidence levels across MoE layers.