Researchers from Hugging Face have introduced Uno, a new class of diffusion-augmented autoregressive (AR) language models designed to overcome the sequential bottleneck of traditional next-token prediction. By decoupling model parameters into AR weights and lightweight diffusion weights, Uno enables parallel generation of multiple tokens while preserving the original AR distribution.
The diffusion weights are trained through a simple Diffusion Distillation phase that adds minimal overhead to existing LLM training pipelines. The accompanying Ψ-Spec sampler guarantees lossless acceleration and supports inference-time scaling at fixed context lengths, eliminating the need for separate draft models used in speculative decoding.
According to the paper, Uno achieves higher throughput than leading speculative-decoding methods at every evaluated batch size, delivering up to 3× speedups over the base AR model. Notably, the 8B Uno model outperforms the 26B DiffusionGemma and the proprietary Mercury 2 across benchmarks in agentic tool use, coding, and long-context reasoning.
In an online discussion, the authors addressed comparisons with the earlier Orthrus framework, clarifying that Uno keeps the architecture unchanged, whereas Orthrus adds diffusion attention heads. They acknowledged similarities and plan to add a discussion in the next revision. Community members also raised questions about wall-clock performance at realistic batch sizes and the breadth of evaluated architectures, which the authors may address in future work.