Nvidia debuts Nemotron 3 with hybrid MoE and Mamba-Transformer (venturebeat.com)

Nvidia has unveiled Nemotron 3, a new version of its AI model architecture that focuses on improved accuracy and reliability for multi-agent systems. The lineup includes three variants: Nemotron 3 Nano with 30 billion parameters for targeted tasks, Nemotron 3 Super at 100 billion parameters for high-accuracy applications, and Nemotron 3 Ultra with approximately 500 billion parameters for complex tasks. Utilizing a hybrid mixture-of-experts (MoE) architecture, Nvidia aims to enhance scalability and efficiency, allowing enterprises greater flexibility and performance in building autonomous systems. The hybrid Mamba-Transformer architecture incorporated in Nemotron 3 enables up to 4x higher token throughput compared to the previous version and significantly reduces inference costs. Remarkably, the larger models implement a "latent MoE" innovation that allows experts to share a common core for efficiency, akin to chefs sharing a kitchen with individual spice racks. In addition, Nvidia is fostering community engagement by launching the NeMo Gym, a reinforcement learning lab for testing models, alongside providing access to research papers, datasets, and sample prompts. With these advancements, Nvidia positions itself to better meet developers' demands for open, efficient, and intelligent AI models.