NVIDIA has unveiled the Nemotron 3 family of open models, which includes the Nano, Super, and Ultra variants, designed to support the growing demand for agentic AI systems across various industries. The family leverages a hybrid mixture-of-experts architecture, enabling high efficiency and accuracy for tasks requiring multiple AI agents.
Nemotron 3 Nano is immediately available, offering a 30-billion-parameter model that activates up to 3 billion parameters per task. It is optimized for low-cost, high-throughput use cases such as debugging, content summarization, and information retrieval. This model achieves up to four times the throughput compared to its predecessor and introduces a 1-million-token context window, significantly improving long-context reasoning.
NVIDIA has just released Nemotron 3 Nano, a ~30B MoE model that scores 52 on the Artificial Analysis Intelligence Index with just ~3B active parameters
— Artificial Analysis (@ArtificialAnlys) December 15, 2025
Hybrid Mamba-Transformer architecture: Nemotron 3 Nano combines the hybrid Mamba-Transformer approach @NVIDIAAI has used on… pic.twitter.com/sStgiDZTYT
The Nemotron 3 Super and Ultra models, with 100 billion and 500 billion parameters respectively, are tailored for complex, multi-agent workflows and are scheduled for release in the first half of 2026. These models utilize NVIDIA's 4-bit NVFP4 training format on Blackwell architecture, reducing memory needs and accelerating training, allowing larger models to run efficiently on existing infrastructure.
NVIDIA, a leader in GPU and AI technology, is providing not only the models but also open training datasets and reinforcement learning libraries to foster transparent and specialized AI agent development. The Nemotron 3 suite supports deployment on a range of platforms, including major cloud services and on-premises NVIDIA-accelerated infrastructure. Early integration by prominent enterprises and startups signals strong industry interest, with independent benchmarking organizations highlighting Nemotron 3's efficiency and accuracy in its class.
