“The Qwen team predicts that leading large language models will scale to 3 trillion, 5 trillion, and potentially over 10 trillion parameters in the future.”