NVIDIA Nemotron is a family of open AI models made by NVIDIA, built mainly for AI agents. Open means that the models and their training data are published on Hugging Face and may be used, changed and deployed commercially under open licenses such as the NVIDIA Open Model License. The family covers reasoning language models, vision-language models, speech, retrieval (document parsing, embeddings and reranking) and safety models.
The speech models do automatic speech recognition, text-to-speech and machine translation. One of them, Nemotron 3 Diarization, performs speaker diarization: it takes audio and outputs which speaker spoke when, for up to eight speakers, on a live stream or on a finished recording, and it can be combined with streaming speech recognition to label each part of a transcript. This model uses the OpenMDW-1.1 license. Nemotron models run on NVIDIA hardware, from RTX PRO workstations and DGX Spark to the cloud.
The models are free to download. NVIDIA also offers free API endpoints for prototyping, and third-party inference providers host Nemotron with usage-based pricing.