First large-scale 3B/8B continuous diffusion LMs match pass@1 and beat pass@k vs masked dLMs
ArashVahdat · x · 2026-10-06
Morteza Mardani and colleagues (retweeted by NVIDIA's Arash Vahdat) present the first large-scale (3B/8B) continuous diffusion language models.
Key results:
- Leveraging smooth continuous trajectories, the models achieve competitive pass@1 and superior pass@k reasoning versus masked diffusion LMs (dLMs)
- Quality degrades gracefully at low NFEs (number of function evaluations)
- The authors frame it as an important step toward single/few-step decoding and efficient RL post-training at scale
More from Models
- Questioning the launch: new model mirrors PrismML scales and kernels without attribution — _xjdr · 2026-10-06
- $3,500 Blackwell Personal AI PC: RTX PRO 4000 Runs Qwen Next at 50-70 tok/s — Jackyhuang · 2026-10-06
- Reddit user: MiniMax H3 and ref mods are "really incredible" — CompleteBed1797 · 2026-10-06
- Reddit users report sudden wave of refusals from Claude with no clear cause — astrorocks · 2026-10-06
- Bought an RTX 5060 for local LLMs — complex tasks scored 2/10 vs 9/10 in the cloud — Tricky-Brother-7 · 2026-10-06
- Full-bandwidth Transformer paper revised: latent feedback nears 1.5x-token performance — _arohan_ · 2026-10-06