Uno: Diffusion-Augmented LLM Claims Lossless Parallel Generation, Beating Speculative Decoding
A research team spanning UIUC, Cornell, Cerebras and others (authors include Eric Xing and Zhengzhong Liu; ssahoo among the announcers) proposes Uno, a diffusion-augmented LLM aiming to fix two weaknesses of diffusion LLMs versus autoregressive (AR) models: lower generation quality and slower large-batch inference. The IFM team also released the companion model K2-Horizon-7B-Uno. Community replications, however, have already disputed some of its headline claims.
Confirmed
- Design: keeps the AR architecture, with each layer holding two sets of weights — standard NTP-trained AR weights plus diffusion-adapted weights, with AR verification
- Inference: a diffusion process draws multiple tokens in parallel from the same distribution as the AR model, so no separate draft model is needed and quality of the replaced model is preserved (authors claim "lossless parallelism"), yielding a claimed lossless 3x speedup
- Claimed performance: an 8B Uno beats 26B DiffusionGemma and Mercury 2; inference speed exceeds EAGLE-3 speculative decoding and all existing diffusion LLMs; faster at large batch sizes
- Release: IFM published K2-Horizon-7B-Uno
Unconfirmed
- Community discussion relayed by @joaogante reports testers questioning Uno's "beats peers" claims as overstated, with replications diverging from the paper; unresolved
Why it matters
- Speculative decoding (e.g. EAGLE-3) needs an extra draft model and may lose quality; if Uno's same-distribution parallel sampling holds up, it addresses both quality and throughput pain points
- It offers a new paradigm for merging diffusion and AR LLMs, but its performance claims are contested and await independent replication
2026-09-04 ~ 2026-09-06 · 6 related posts
Primary sources
- Diffusion-augmented LLM Uno: 8B beats 26B DiffusionGemma with 3x lossless speedup — iScienceLuvr ·
- Uno pairs AR weights with diffusion LoRA adapters, beating EAGLE-3 and all diffusion LLMs — HongyiWang10 ·
- Uno: diffusion-augmented LLMs enable lossless parallel token generation, beating speculative decoding — dair_ai ·
- [source] Diffusion-augmented LLM Uno: 8B beats 26B DiffusionGemma with 3x lossless speedup — iScienceLuvr · 2026-09-04
- Uno adds diffusion weights to AR LLMs, beating EAGLE-3 and all diffusion LLMs — JFPuget · 2026-09-04
- Uno hybrid diffusion LLM claims 'beats all', but latency-quality is Pareto dominated — joao_gante · 2026-09-04
- [source] Uno pairs AR weights with diffusion LoRA adapters, beating EAGLE-3 and all diffusion LLMs — HongyiWang10 · 2026-09-05
- [source] Uno: diffusion-augmented LLMs enable lossless parallel token generation, beating speculative decoding — dair_ai · 2026-09-05
- Diffusion adapters give lossless 3x LLM speedups with no draft model, Uno paper shows — kastnerkyle · 2026-09-06