NVIDIA announces the Nemotron 3 series, a compact, high-performance open AI model

On December 15, 2025, NVIDIA announced the Nemotron 3 family of open AI models, data, and libraries designed to enhance specialized agent AI development and bring transparency and efficiency to the entire AI industry. By activating only the parts that are needed, Nemotron 3 enables efficient operation of multi-agent systems while reducing costs.

NVIDIA Debuts Nemotron 3 Family of Open Models | NVIDIA Newsroom
https://nvidianews.nvidia.com/news/nvidia-debuts-nemotron-3-family-of-open-models/?ncid=so-twit-561360

The Nemotron 3 family is an MoE model that combines multiple specialized agents, and is available in three sizes: the compact Nemotron 3 Nano with 30 billion parameters; the high-precision Nemotron 3 Super with approximately 100 billion parameters that can activate up to 10 billion parameters per token; and the large-scale Nemotron 3 Ultra with approximately 500 billion parameters that can activate up to 50 billion parameters per token.

The Nemotron family achieves superior accuracy through advanced reinforcement learning techniques that enable simultaneous training of large-scale multiple environments, while at the same time demonstrating extremely high token generation capabilities that are best-in-class for large-scale multi-agent systems through its groundbreaking hybrid MoE architecture. According to NVIDIA, Nemotron 3 Nano achieves four times higher throughput than the Nemotron Nano 2 , which will be released in August 2025.

Artificial Analysis, an AI model analysis company, rated Nemotron 3 Nano as 'the most open and efficient model in its class,' reporting that it achieved top-class accuracy across coding, inference, and agent benchmarks. In pre-release tests, it recorded an output speed of approximately 380 tokens per second, and stated that 'it focuses not only on intelligence but also on efficiency, achieving a compelling trade-off between speed and performance.'

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

Hybrid Mamba-Transformer architecture: Nemotron 3 Nano combines the hybrid Mamba-Transformer approach @NVIDIAAI has used on… pic.twitter.com/sStgiDZTYT

— Artificial Analysis (@ArtificialAnlys) December 15, 2025

Nemotron 3 Super and Ultra also feature efficiencies that significantly reduce memory requirements and speed up training. This enables larger models to be trained on existing infrastructure without sacrificing accuracy compared to high-precision formats. Nemotron 3 Super and Ultra run on NVIDIA's Blackwell architecture and use the NVFP4 training format, reducing memory requirements and enabling large-scale training on existing infrastructure.

Nemotron supports NVIDIA's broader AI efforts, embracing an open, transparent and efficient model that enables organizations around the world to build AI systems aligned with their own data, regulations and values. 'Open innovation is the foundation of advances in AI. With Nemotron, we're transforming advanced AI into an open platform, giving developers the transparency and efficiency they need to build large-scale agent systems,' said NVIDIA CEO Jensen Huang.

Nemotron 3 Nano is available for download on multiple platforms, including Hugging Face. The dataset released for Nemotron includes pre-training, post-training, and reinforcement learning data totaling 3 trillion tokens, and the libraries are also available as open source. Nemotron 3 Super and Ultra are scheduled for release in the first half of 2026.

nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8 · Hugging Face
https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-FP8