The Uno method utilizes lightweight diffusion weights grafted onto existing autoregressive models. This architecture allows the model to generate candidate token blocks in parallel, while the base model serves as a quality referee to ensure the output matches the original distribution.
Lead author Subham Sekhar Sahoo and his team introduced Psi-Spec samplers. These diffusion-based adapters enable the simultaneous generation of multiple candidate tokens, maintaining the original autoregressive weights for final verification.
The Uno-Qwen 8B variant, based on the Qwen 8 billion parameter model, demonstrates a 2.5x speedup at batch size 1 and approximately 1.6x at larger batch sizes. The overall system throughput can reach up to 3x compared to standard autoregressive models.
The 8B parameter Uno model outperforms the 26B parameter DiffusionGemma across coding, agentic, and long-context benchmarks. It also claims lower peak memory usage than other acceleration methods, including speculative decoders like EAGLE-3 and DFlash.
The team has released the code and checkpoints on GitHub in the ifm-ai/uno repository and model weights on Hugging Face (s-sahoo/uno-qwen3-8B). The release supports inference via Nano-vLLM with linear and tree samplers.
The technique was detailed in the preprint paper 'Unlocking Lossless Speedups in LLMs via Discrete Diffusion' published on September 3, 2026. Independent verification of the performance claims is currently pending.