Nvidia Nemotron 3 Ultra (research.nvidia.com)

Nvidia has unveiled its most powerful model to date, the Nemotron 3 Ultra, which boasts an impressive 550 billion total parameters and 55 billion active parameters. This model represents the pinnacle of the Nemotron 3 series and incorporates several advanced features, including a Mixture-of-Experts Hybrid Mamba-Attention architecture and LatentMoE for enhanced accuracy. Its innovative inference mechanisms leverage native speculative decoding, enabling faster performance while maintaining reasoning budget control. The model shows remarkable improvements in inference throughput, achieving up to 5.9 times higher speed compared to previous state-of-the-art models. The Nemotron 3 Ultra also excels in context handling, supporting input lengths of up to 1 million tokens and outperforming competitors on benchmark tasks. Its open-source release includes pre-trained and post-trained checkpoints, along with diverse datasets aimed at enhancing various capabilities such as legal understanding and factual recall. By sharing these resources, Nvidia not only fosters innovation within the AI/ML community but also sets a new standard for the performance of large language models, emphasizing both speed and versatility in generative tasks.