Try llama.cpp - it usually excels with these MoE models imho.