“The Qwen-3 30B MoE model, which activates 3 billion parameters, can outperform the Qwen-2 32B dense model in some tasks.”