Uno adds diffusion weights to AR LLMs, beating EAGLE-3 and all diffusion LLMs
JFPuget · x · 2026-09-04
New method Uno tackles diffusion LLMs' two key limitations vs AR models: lower quality and slower inference at large batch sizes.
- Keeps the AR architecture of LLMs, with each layer holding two weight sets: AR weights and diffusion weights
- The diffusion weights enable lossless parallel sampling from the AR distribution
- Faster than all speculative decoding methods, including DFlash and EAGLE-3
- Outperforms every existing diffusion LLM: Mercury 2, Diffusion Gemma, and Llada
Paper, models, and code are all linked.
More from Models
- Fable 5.5 delayed: 'Stoptober' meme says no one dares ship a SOTA model now — patience_cave · 2026-10-21
- Decision Models' routing classifier hits AntSeed, cutting inference cost to 24% of always-Opus — airesearch12 · 2026-10-21
- Developer flags anomalous Claude Projects usage: 98% weekly limit consumed at 1% thread share — 0xkarasy · 2026-10-20
- DeepSeek hands inference providers 90%+ margins, pressuring Moonshot and Zhipu hosting deals — teortaxesTex · 2026-10-20
- HelixWorld world model tested: 70GB VRAM per BF16 GPU run, consumer cards need not apply — cocktailpeanut · 2026-10-20
- Code-only Paris at golden hour: Claude's render embarrasses ChatGPT in same-prompt test — EXM7777 · 2026-10-20