NVIDIA Releases Nemotron-Labs-Sigma-distill-8B-nfe2 Model
NVIDIA has launched the Nemotron-Labs-Sigma-distill-8B-nfe2, an 8 billion-parameter continuous diffusion language model.
NVIDIA has launched the Nemotron-Labs-Sigma-distill-8B-nfe2, an 8 billion-parameter continuous diffusion language model designed for fast few-step generation. This model, distilled using Phased Distribution Matching Distillation (Phased DMD), is part of the Nemotron-Labs-Sigma architecture family and is tailored for latency-sensitive generation research, speculative and parallel decoding studies, and deployment-oriented experimentation. It features a maximum context length of 4,096 tokens and generates text in 32-token blocks using exactly 2 denoising function evaluations (NFE) plus one final readout. The model is available for commercial or non-commercial use under the NVIDIA Nemotron Open Model License, and it can be run offline on a user’s own hardware as NVIDIA hosts no service. According to the model card, the model achieves faster training and inference times when optimized to run on NVIDIA GPU-accelerated systems. It requires PyTorch (>= 2.4) with Hugging Face Transformers (>= 4.55) and a CUDA-capable NVIDIA GPU for execution, with CPU inference not supported due to numerical degradation.