Uno (Diffusion-Augmented LLMs)

"Unlocking Lossless Speedups in LLMs via Discrete Diffusion." IFM at MBZUAI, with Cerebras co-authors, introduces diffusion-augmented LLMs: a model that defines an ordinary autoregressive distribution but uses diffusion to draw several tokens in parallel from it. The parameters split into AR weights trained with next-token prediction and lightweight diffusion weights (a conditional LoRA) trained by diffusion distillation; the Ψ-Spec samplers then give provably lossless multi-token generation with no separate draft model, reaching up to 3× the base model's speed at every batch size. Released as Uno adapters for K2 Horizon 0.9B and 7B and for Qwen3-8B, under Apache 2.0.