“The target cache generated during DeepSpec's data preparation stage can be approximately 38 TB for the default Qwen/Qwen3-4B setting.”