No, they still operate on tokens, it's just the difference in how text is sampled. It's done through iterative denoising and in parallel, across multiple token positions.