What's this about?
Uno keeps an autoregressive model's distribution but adds lightweight diffusion weights trained to generate multiple tokens in parallel. The authors report up to 3× throughput over the base model without a separate speculative-decoding draft model.
What exactly is “lossless acceleration”?
It means sampling faster without changing the probability distribution and therefore expected quality of the underlying model. It is closer to parallelising the same calculation than replacing it with an approximation.
Why is this worth noting?
The paper claims an 8B Uno beats larger diffusion baselines across tool use, coding and long-context tests. If reproduced, it blurs the boundary between left-to-right and diffusion generation rather than asking developers to choose one architecture.