NVIDIA NeMo is an open-source framework from NVIDIA for building generative AI models. First released in 2019 as a toolkit for conversational AI (speech recognition, natural language processing, and speech synthesis), it has grown into a platform that covers training, fine-tuning, and deploying large language and multimodal models, with an emphasis on large-scale distributed training on NVIDIA GPUs.
It consists of purpose-specific libraries and microservices, including NeMo Curator for data curation, NeMo Guardrails for controlling conversational behavior, and NeMo Retriever, a set of document parsing, embedding, and reranking microservices for retrieval-augmented generation. Libraries for developing and operating AI agents, such as NeMo Relay for recording agent execution traces and NeMo Switchyard for routing requests among models, are also published under the NeMo name.