November 20, 2026 · MarkTechPost
Nvidia releases open Nemotron-Labs-Sigma diffusion language models
Nvidia released Nemotron-Labs-Sigma, 14 open checkpoints at 3 billion and 8 billion parameters. The models generate text by denoising in a learned 16-dimensional embedding space rather than predicting one token at a time. Distilled few-step samplers decode up to 4.2x faster than autoregressive decoding.
Why it matters: Diffusion is the main serious challenger to token-by-token generation, and its pitch is speed. Open checkpoints let researchers test whether the latency gains hold without large quality losses. Paired with Nvidia's reported Inception acquisition, it suggests Nvidia sees diffusion as a strategic bet for its open Nemotron family, not a side experiment.