The authors propose diffusion-augmented LLMs, called Uno, that combine autoregressive and diffusion approaches. The model decouples parameters into AR weights trained with next-token prediction and lightweight diffusion weights learned via Diffusion Distillation. This allows generating multiple tokens in parallel without a separate draft model.
Uno achieves higher throughput than leading methods at all batch sizes and delivers up to 3x speedups over the base AR model, even at the largest batch size. The 8B Uno model outperforms larger models like 26B DiffusionGemma and Mercury 2 on agentic tool use, coding, and long-context reasoning.
This is interesting because it provides lossless acceleration for LLM inference without requiring additional draft models or sacrificing quality, and it can be built from existing open-weight models.