“The Qwen 30B MoE model has an inference speed comparable to a 7 billion parameter dense model.”