NVIDIA releases Nemotron-Labs Diffusion models

nvidia released nemotron-labs diffusion, a family of open models that shift from token-by-token generation to parallel generation with iterative refinement. the release includes 3b, 8b, and 14b parameter text models under a commercial license and an 8b vision-language model for research. the architecture supports autoregressive, diffusion, and self-speculation modes within a single model to optimize gpu utilization and speed. benchmarks indicate the 8b model achieves 1.2% higher average accuracy than qwen3 8b while offering faster inference.