Large Language Models are powerful but painfully slow, forced to generate text one token at a time like a reader tracing each word with a finger. This sequential bottleneck limits their speed, but researchers have now introduced a method that breaks this chain without sacrificing accuracy. By combining the reliability of autoregressive models with the parallel power of diffusion, they have created a new class of models that generate multiple tokens simultaneously.

The core innovation lies in decoupling the model into two parts. The main autoregressive weights handle the standard next-token prediction, while lightweight diffusion weights are trained to generate several tokens in parallel. This diffusion training is added via a simple distillation process that adds negligible overhead to existing pipelines. The result is a system that draws from the same distribution as the original model but does so in a burst, effectively parallelizing the generation process.

This approach solves a major conflict in current acceleration techniques. Unlike speculative decoding, it requires no separate draft model to guess the next steps. Unlike other diffusion-based LLMs, it does not degrade the quality of the underlying autoregressive model. The researchers call these new models Uno and introduce a sampler family that enables lossless acceleration. This means you get faster speeds without the typical trade-off of reduced coherence or accuracy.

The performance gains are substantial. Uno models achieve higher throughput than leading speculative-decoding methods across all batch sizes and deliver up to three times the speed of the base autoregressive model. Notably, their eight-billion-parameter Uno model outperforms larger twenty-six-billion-parameter diffusion models and proprietary systems in coding, tool use, and reasoning benchmarks. This breakthrough allows developers to build faster, more efficient AI systems by simply augmenting existing open-weight models with these new diffusion capabilities.