The Qwen 3.6-35B A3B language model contains 40 layers, with each layer comprising 256 experts containing 512 neurons each