To be very clear: as good as it is, Z-Image is NOT multi-modal or auto-regressive, there is NO difference whatsoever in how it uses Qwen relative to how other models use T5 / Mistral / etc. It DOES NOT "think" about your prompt and it never will. It is a standard diffusion model in all ways.