"Batch output to multiply speed gains A core reason diffusion language models are dramatically faster than autoregressive ones is architectural: 'in the autoregressive world if you want to generate a thousand tokens you need a thousand neural network evaluations,' whereas in a diffusion model 'the neural network can output many tokens at every step.' This means speed gains compound — fewer steps times more tokens per step. Any pipeline that processes items sequentially when they could be batched is leaving the same kind of speed on the table. Source: Arash Vahdat, Latent Space Podcast, Diffusion LLMs with Inception AI"