NVIDIA team releases Sigma: first large-scale continuous diffusion LM with 3B/8B checkpoints
LucaAmb · x · 2026-11-17
The NVIDIA Nemotron Labs team has open-sourced 3B/8B checkpoints for Sigma on Hugging Face — the first large-scale continuous diffusion language model (dLM) trained with likelihood.
- Matches discrete dLMs and autoregressive models on math and coding benchmarks
- Enables meaningful study of inference-time interventions, a key research angle for the diffusion-LM paradigm
- Jointly led by zhihanyang, Wei Guo, Jean-Marie Lemercier and Simon Welker, with NVIDIA researchers including Arash Vahdat on the team
Checkpoints are available in the Nemotron-Labs-Sigma collection for community reproduction and follow-up research.
More from Research
- Grant project seeks $5K–$20K to study how agent personalities shape their societies in OHOL — sebkrier · 2026-11-17
- A Pharma Agent That Flags Developability Risks on Every Sequence Registration — nlarusstone · 2026-11-17
- After CoRL, roboticists debate where agentic robotics breaks: sim, dexterity, repeatability — k7agar · 2026-11-17
- Polyak–Lojasiewicz paper 10-year retrospective published, wins 2026 KDD Test of Time award — jasondeanlee · 2026-11-17
- Microsoft Details Maia Dataflow Architecture: 4719 Tflop/s fp8, Over 99% of Peak — thoefler · 2026-11-17
- Ternary Moral Logic: A Proposal for Hardware-Enforced AI Pause and Accountability — MoralLogs · 2026-11-17