Uno: diffusion-augmented LLMs enable lossless parallel token generation, beating speculative decoding
dair_ai · x · 2026-09-05
dairai highlights Uno, a new class of diffusion-augmented LLMs. Instead of speculative decoding's separate draft model, Uno draws multiple tokens in parallel from the autoregressive model's own distribution via diffusion.
Key points:
- Parameters split in two: AR weights train with standard next-token prediction, plus a lightweight diffusion head added by a short distillation phase with negligible overhead.
- Because sampling comes from the AR distribution itself, the speedup is lossless, and existing open-weight AR models can be upgraded without retraining.
- The author reports Uno outperforms leading speculative decoding methods.
More from Models
- Space Bunny Alpha revealed as MiniMax M3.1 Flash: 1M context, budget-friendly coding — ZabihullahAtal · 2026-10-21
- 'Opus 5.5 Medium' Spotted in Rork's Model Picker, Hinting at New Anthropic Model — rudrank · 2026-10-21
- RSI-Bench Clears First Task Set Through Reward-Hacking Detection Pipeline — SRSchmidgall · 2026-10-21
- Dev complains Anthropic's bio safeguards trigger just from reading a bio paper — robleclerc · 2026-10-21
- Redditor claims uncensored local DeepSeek build answers virus-writing prompt with zero hesitation — Only_Light_9817 · 2026-10-21
- Interconnects Artifacts 25: 28 new models, DeepSeek V4.1 Flash, and US-China regulatory controls — natolambert · 2026-10-21