Nvidia Nemotron 3 Ultra 550B-A55B BF16 (huggingface.co)

Nvidia has unveiled its Nemotron 3 Ultra 550B-A55B BF16 model, a groundbreaking large language model featuring 550 billion total parameters and 55 billion active parameters. This model utilizes a hybrid Latent Mixture-of-Experts architecture, combining Mamba-2, Mixture of Experts (MoE), and Attention layers, enhanced with Multi-Token Prediction (MTP) for optimized text generation. Designed for intensive workloads, it supports a massive context length of up to 1 million tokens and is capable of handling complex reasoning tasks, multilingual interactions, and high-stakes information retrieval. The significance of this release lies in its potential to advance the development of AI agents and chatbots, particularly those requiring high-quality reasoning and long-context analytical abilities. Developers can leverage this model for sophisticated tasks in fields like coding, math, and science, benefitting from its configurable reasoning mode. By employing advanced techniques like Multi-Domain On-Policy Distillation (MOPD) and asynchronous reinforcement learning, the model ensures robust performance across various domains. With support for multiple languages, this model opens new avenues for cross-lingual applications, solidifying Nvidia's position as a leader in the AI/ML landscape.