“The MoE implementation in Qwen 3 Next is sparser than previous versions, using 512 experts and activating 10 per token.”