Uno bolts diffusion adapters onto an autoregressive LLM for 3x lossless speedup, and the weights are already up.

The model splits 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. No separate draft model, unlike speculative decoding. No quality loss, unlike diffusion LLMs. Up to 3x over the base AR model and higher throughput than leading speculative decoding at every batch size, with an 8B Uno beating the 26B DiffusionGemma. IFM/K2-Horizon-7B-Uno and IFM/K2-Horizon-0.9B-Uno are live on Hugging Face as Apache-2.0 conditional-LoRA adapters. arXiv 2609.04010