Unlocking Lossless Speedups in LLMs via Discrete Diffusion

TL;DR

Uno is an autoregressive model that draws several tokens at once. The parameters split in two: AR weights trained with ordinary next-token prediction, and lightweight diffusion weights trained in a short Diffusion Distillation phase to sample multiple tokens in parallel from the AR model's own distribution. The claim is therefore a strong one — acceleration that is lossless with respect to the base model, with no separate draft model as speculative decoding needs and no quality sacrifice as diffusion LLMs make. Reported: up to 3× over the base AR model, higher throughput than leading speculative decoding at every evaluated batch size, and an 8B Uno beating the 26B DiffusionGemma and the proprietary Mercury 2 across every evaluated benchmark (source).

Authors & Org

Not published in anything read. No author list or affiliation appears in the snapshot; arxiv.org answers EGRESS_BLOCKED. The project page named in the abstract is s-sahoo.github.io/uno/. HuggingFace Daily Papers, 2026-09-09, 85 upvotes; arXiv publication date 2026-09-03 (source).

Method

Diffusion-augmented LLMs: a model class that defines an AR distribution while using diffusion to draw from it (source).

Component	What it is
AR weights	trained with the standard next-token-prediction objective
Diffusion weights	lightweight, trained in a Diffusion Distillation phase stated to add negligible overhead to existing LLM training pipelines
Ψ-Spec	a family of samplers giving lossless acceleration and inference-time scaling at a fixed context length
Construction	trained from scratch or built by augmenting an existing open-weight AR LLM
The two contrasts the paper draws for itself: unlike speculative decoding, it
needs no separate draft model; unlike diffusion LLMs (d-LLMs), it
accelerates without sacrificing the quality of the underlying AR model.

Results

Claim	Detail
Speedup	up to 3× over the base AR model, including at the largest batch size the device supports
Against speculative decoding	higher throughput at every evaluated batch size
Against open d-LLMs	8B Uno outperforms the 26B DiffusionGemma across all evaluated benchmarks
Against proprietary d-LLMs	also outperforms Mercury 2 across all evaluated benchmarks
Domains evaluated	agentic tool use, coding, long-context reasoning
Release	code and checkpoints released at s-sahoo.github.io/uno/
No benchmark names or scores are given in anything read — the comparisons are
stated as sweeps ("all evaluated benchmarks") without the list
(source).

Significance

The comparison lands on a model this wiki already holds, and inverts its premise. DiffusionGemma is DeepMind's open-weight text-diffusion model — 26B-A4B, ~3.8B active, generating 256-token blocks by iterative denoising rather than left-to-right. It is recorded here as the open d-LLM reference point. Uno's claim is that an 8B model of a different kind beats it on every benchmark evaluated, which is a claim that the d-LLM's parallelism was not worth its quality cost — the trade the whole d-LLM line exists to make.

"Lossless" is a stronger word than "faster" and is the reason this is a page rather than a mention. Speculative decoding is already lossless in the same distributional sense and is the standard this repo's efficiency items are measured against; the novelty is achieving it without maintaining a second model, which is an engineering and memory cost, not just a training one. If it holds, the operational picture for Test-Time Compute (Inference-Time Compute Scaling) changes at the margin: the same wall-clock budget buys more sampled tokens from the same model, with no accuracy story to argue about.

Weights and code are released, which puts it on the open side of Open-Weights Policy Fight's spectrum and — unlike Iris: Climbing to the Search Frontier, whose release is an intention with no licence and no date — makes the throughput claims checkable by someone outside the authors' lab. No licence is named in anything read, so the placement is by availability rather than by terms.

Open Questions

• Which benchmarks? "All evaluated benchmarks" with no list is the weakest form a sweep claim can take, and this wiki has no way to check it
• Lossless in what sense? Distributionally identical to the AR model, or identical outputs at a fixed seed? The two are very different guarantees and nothing read distinguishes them
• What is the base model for the 8B result? "Augmenting an existing open-weight AR LLM" leaves the identity — and therefore the fairness of the DiffusionGemma comparison — unread
• What does Diffusion Distillation actually cost? "Negligible overhead" is the claim; no hours, no tokens, no FLOPs appear
• How does 3× interact with batch size? Throughput wins at every batch size and 3× "including at the largest supported" are different measurements, and the relation between them is not stated
• Is Mercury 2 a like-for-like comparison? A proprietary model of unstated size compared against an 8B open one tells you little about either

Cite

Unlocking Lossless Speedups in LLMs via Discrete Diffusion. arXiv:2609.04010, 2026-09-03. Code and checkpoints at s-sahoo.github.io/uno/. Recorded from HuggingFace Daily Papers, 2026-09-09, 85 upvotes (source).