A fully ready Qwen3-4B-Thinking-2507 build. Compared to vanilla Qwen3-4B, it delivers noticeably better prompt adherence with Z-Image models and avoids common wording misinterpretations. Highly recommended for both inference and Z-Image LoRA training.

Path: ComfyUI_windows_portable\ComfyUI\models\text_encoders\

training config for AiToolKit with Qwen3-4B-Thinking-2507 text encoder Can be found here.

full training thread can be found here