NVIDIA Releases Nemotron-Labs-Sigma-distill-3B-nfe8 Model

NVIDIA has launched the Nemotron-Labs-Sigma-distill-3B-nfe8, a 3 billion-parameter continuous diffusion language model.

The Nemotron-Labs-Sigma-distill-3B-nfe8 model is designed for fast few-step text generation with a maximum context length of 4,096 tokens. It is a distilled version of a 3B teacher model from the same architecture family, utilizing Phased Distribution Matching Distillation (Phased DMD). The model generates text in 32-token blocks using exactly 8 denoising function evaluations (NFE) plus one final readout. It is available for both commercial and non-commercial use under the NVIDIA Nemotron Open Model License. The model can be run locally on a user’s hardware, specifically requiring a CUDA-capable NVIDIA GPU, as CPU inference is not supported due to numerical degradation. NVIDIA claims that the model achieves faster training and inference times when leveraging NVIDIA’s GPU-accelerated systems and software frameworks. The model is integrated with PyTorch (>= 2.4) and Hugging Face Transformers (>= 4.55), and it is optimized to run on NVIDIA’s Ampere, Blackwell, Hopper, and Lovelace microarchitectures.