Uno Bolts Diffusion Adapters Onto an Autoregressive LLM for 3x Lossless Speedup, and the Weights Already Shipped

arXiv 2609.04010·high signalXBlueskyLinkedInCopy link

The Uno paper splits an LLM into standard AR weights trained with next-token prediction plus lightweight diffusion weights learned in a short distillation phase, letting diffusion draw multiple tokens in parallel from the AR model's own distribution, so unlike speculative decoding there is no separate draft model and unlike diffusion LLMs there is no quality loss. Uno reports up to 3x speedup over the base AR model and higher throughput than leading speculative-decoding methods at every batch size, with an 8B Uno beating the 26B DiffusionGemma. This is not paper-only: IFM/K2-Horizon-7B-Uno and IFM/K2-Horizon-0.9B-Uno are live on HuggingFace as Apache-2.0 conditional-LoRA diffusion adapters over the K2-Horizon base weights, with code at github.com/ifm-ai/uno created 2026-09-03.