Nemotron

NVIDIA Nemotron is a family of open AI models built for long-running agentic workflows. It covers several workload types: reasoning models in Nano, Super and Ultra sizes, multimodal vision models (the site names Nemotron 3.5 Super VL), speech models (ASR, TTS and neural machine translation), retrieval models (embeddings, reranking, multimodal document parsing) and safety models for content moderation and jailbreak protection.\n\nNVIDIA describes the models as using a hybrid mixture-of-experts architecture tuned for throughput. It says models and training data are published openly on Hugging Face. The multimodal models are positioned for document intelligence, computer-use agents, and video and audio understanding. Research models are also offered for experimentation.\n\nThe models can be downloaded from Hugging Face or run as NVIDIA NIM microservices. NVIDIA also offers free prototyping API endpoints, and the models work with NeMo (customization) and NVIDIA Blueprints (reference workflows). Hardware support named on the site includes RTX PRO and DGX Spark. Licensing differs per model and is not detailed in the provided material.