A lightweight diffusion adapter lets large language models generate multiple tokens in parallel without sacrificing output quality
Getting a large language model to think faster usually means making it think worse. A new method called Uno claims to break that tradeoff, delivering up to 3x throughput improvements over standard autoregressive models while preserving the quality of every generated token.
The technique, detailed in a paper titled “Unlocking Lossless Speedups in LLMs via Discrete Diffusion” (arXiv:2609.04010) published on September 3, 2026, works by grafting lightweight diffusion weights onto an existing autoregressive model. Instead of replacing the original architecture, it augments it, letting the model generate candidate token blocks in parallel while the base model acts as a quality referee.
How Uno actually works #
Uno’s lead author Subham Sekhar Sahoo and collaborators introduce what they call Psi-Spec samplers, which use diffusion-based adapters to generate multiple candidate tokens simultaneously. The autoregressive weights stay intact and handle final verification, so the output quality matches what the base model would have produced on its own.
The key insight is that diffusion models are naturally parallel. While autoregressive generation is inherently sequential (token B depends on token A), diffusion can refine a noisy block of tokens all at once. Uno exploits this property without throwing away the autoregressive model’s strengths.
The result is a system that uses the fewest additional parameters of any comparable acceleration method while delivering higher aggregate throughput than leading speculative decoders like EAGLE-3 and DFlash.
The numbers #
The Uno-Qwen 8B variant, built on top of Qwen’s 8 billion parameter model, achieves roughly 2.5x speedup at batch size 1. At larger batch sizes, the improvement narrows to about 1.6x. The ceiling is up to 3x system throughput compared to the base autoregressive model.
The 8B parameter Uno model outperforms DiffusionGemma, a 26B parameter model, across agentic, coding, and long-context benchmarks. Uno also claims lower peak memory usage than competing acceleration methods.
Open source, but unverified #
The team has released code and checkpoints on GitHub under the repository ifm-ai/uno, with model weights available on Hugging Face (s-sahoo/uno-qwen3-8B). The release includes training recipes and supports inference through Nano-vLLM with both linear and tree samplers.
As of early September 2026, independent verification of Uno’s claimed performance is still pending. The paper is a preprint, not yet peer-reviewed. Uno’s architecture, where the autoregressive model has final say over token selection, is designed to guarantee that outputs match the base model’s distribution exactly.
Why this matters for AI infrastructure #
A 2.5x throughput improvement, if it holds up in production environments, means serving the same number of requests with roughly 40% fewer GPUs. The fact that Uno works as an adapter rather than a replacement means organizations that have already fine-tuned and deployed autoregressive models don’t need to start from scratch—they can bolt Uno’s diffusion weights onto their existing infrastructure.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our