“The Qwen 3 Next model architecture uses a hybrid attention mechanism where for every four layers, three are a Gated Delta Network (a type of linear attention) and one is a full attention layer.”