They make it sound like they improve Qwen by reading raw weights. They don't. In fact the experiments that use raw weights have very little to do with the experiments that work on larger models.