NVIDIA Corp. (NASDAQ: NVDA) launched Nemotron 3 Super, a 120-billion-parameter artificial intelligence model designed for multi-agent AI systems. The model uses 12 billion active parameters and features a 1-million-token context window.

The model employs a hybrid mixture-of-experts architecture combining Mamba and transformer layers. NVIDIA states the design delivers up to 5x higher throughput and up to 2x higher accuracy compared to the previous Nemotron Super model.

Companies integrating the model include Perplexity for search functionality, software development firms CodeRabbit, Factory and Greptile for AI agents, and life sciences organizations Edison Scientific and Lila Sciences. Enterprise software companies Amdocs, Palantir, Cadence, Dassault Systèmes and Siemens plan to deploy the model for workflow automation.

NVIDIA released the model with open weights under a permissive license. The company published training methodology including over 10 trillion tokens of datasets, 15 reinforcement learning environments and evaluation procedures. The model was trained on synthetic data generated using reasoning models.

The model runs on NVIDIA's Blackwell platform using NVFP4 precision, which the company states cuts memory requirements and increases inference speed up to 4x faster than FP8 on NVIDIA Hopper systems.

Nemotron 3 Super is available through build.nvidia.com, Perplexity, OpenRouter and Hugging Face. Cloud service providers including Google Cloud's Vertex AI and Oracle Cloud Infrastructure offer access, with Amazon Web Services and Microsoft Azure availability planned.

Dell Technologies and HPE are integrating the model into their enterprise platforms. Additional deployment partners include CoreWeave, Crusoe, Nebius, Together AI, Baseten, CloudFlare, DeepInfra and Fireworks AI.