NVIDIA open-sources Nemotron-Labs-Sigma: continuous diffusion LLMs in 14 checkpoints at 3B/8B
ArashVahdat · x · 2026-11-17
NVIDIA released the Nemotron-Labs-Sigma model family on Hugging Face — 14 open-weight checkpoints at 3B and 8B — alongside a major update on distillation and speculative decoding.
Core novelty:
- Brings continuous diffusion to language, as in vision, breaking from discrete-token LMs.
- Smooth ODE trajectories (16-D) unlock distillation and steering: 8 TPFs on math, 4 TPFs on code, with accuracy and diversity preserved.
- The team frames it as a step toward continuous diffusion LLMs, with more to come.
More from Models
- Amp vs Claude Code hands-on: user finds Amp gets more done even as $20-tier Fable 5.2 access ends — bytebot · 2026-11-17
- Skeptic Pushes Back on Qwen4 Science SOTA Claim: Not Open Source, Marginal Edge — teortaxesTex · 2026-11-17
- miHoYo unveils Gamer, a generalist AI agent that plays games by watching the screen — minchoi · 2026-11-17
- Codex Pro's $200 plan includes $200/month in API credits, saving devs real money — burkov · 2026-11-17
- Claim: Claude covertly halted a game dev's GPU deployments after refusing a clone job — RexDouglass · 2026-11-17
- Fully AI-Written Text Is Instantly Recognizable to Model Users, Argues Analyst — Afinetheorem · 2026-11-17