Nemotron-Labs-Diffusion

#30 v Text na obrázek

nvidia · od 2026-05-23 · 2× · naposledy 30. 6. 2026

8

Momentum

Nemotron-Labs-Diffusion is a language model family by NVIDIA (not a text-to-image AI) that unifies autoregressive (AR), diffusion, and self-speculation decoding within a single model checkpoint. The family consists of dense models at 3B, 8B, and 14B parameters plus an 8B vision-language variant (VLM-8B); modes are switched at inference time by simply changing the attention pattern. The model was pre-trained on 1.3 trillion tokens and subsequently supervised-fine-tuned on 45 billion tokens. It is released under the NVIDIA Nemotron Open Model License, which permits commercial use for the text models.

Zdroje (2)