Researchers have introduced Uno, a new class of diffusion-augmented Large Language Models designed to overcome the sequential bottlenecks of traditional autoregressive (AR) next-token prediction. Unlike standard models that generate tokens one by one, Uno utilizes discrete diffusion to sample multiple tokens in parallel from an AR distribution. The architecture decouples parameters into two distinct sets: standard AR weights trained via next-token prediction and lightweight diffusion weights learned through a "Diffusion Distillation" phase. This approach avoids the need for a separate draft model required by speculative decoding and preserves the original AR model's quality, unlike previous diffusion-only LLMs. The authors implemented $\Psi$-Spec, a family of samplers that enables lossless acceleration and inference-time scaling. Experimental results demonstrate that Uno achieves higher throughput than current speculative-decoding methods across all evaluated batch sizes, delivering up to a $3\times$ speedup over the base AR model. Furthermore, the 8B Uno model was shown to outperform both the 26B DiffusionGemma and the proprietary Mercury 2 model across critical benchmarks, including coding, long-context reasoning, and agentic tool use. This method can be applied to existing open-weight LLMs or used to train models from scratch, providing a scalable path toward faster, high-fidelity inference.