Uno 8B diffusion-augmented LLM claims 3× lossless speedup

An 8B model called Uno claims to match or beat both a 26B open-weight diffusion LLM and a proprietary diffusion LLM across agentic tool use, coding, and long-context reasoning benchmarks, while delivering, according to its authors, up to 3× the throughput of the underlying autoregressive model it augments.

The preprint on arXiv frames Uno as a new class of "diffusion-augmented LLMs" that decouples parameters into AR weights trained on next-token prediction and lightweight diffusion weights trained to generate multiple tokens simultaneously from the same AR distribution. The mechanism is a "Diffusion Distillation phase that adds negligible overhead to existing LLM training pipelines," paired with a sampler family the authors call Ψ-Spec, which they say enables "lossless acceleration and inference-time scaling at a fixed context length."

On the head-to-head with existing acceleration tricks, the abstract is direct: "Unlike speculative decoding, our method requires no separate draft model," and unlike existing diffusion LLMs, "it accelerates generation without sacrificing the quality of the underlying AR model." The authors report Uno "achieves higher throughput than leading speculative-decoding methods at every evaluated batch size" and delivers up to 3× over the base AR model, "including at the largest batch size supported by the device."

The headline quality claim is narrower and easier to check: "our 8B Uno model outperforms the leading open d-LLM, the 26B DiffusionGemma, and the proprietary Mercury 2 across all evaluated benchmarks in agentic tool use, coding, and long-context reasoning." The abstract does not break out per-benchmark scores, name the open-weight AR checkpoints the authors augmented to build the 8B Uno, or specify the hardware and batch sizes behind the 3× figure.