NVIDIA Nemotron-3-Super-120B-A12B-BF16: A Hybrid LatentMoE for Advanced Agentic AI

NVIDIA Nemotron-3-Super-120B-A12B-BF16 is a powerful large language model developed by NVIDIA, featuring a unique Latent Mixture-of-Experts (LatentMoE) architecture. This hybrid design integrates Mamba-2, MoE, and Attention layers, enhanced with Multi-Token Prediction (MTP) for superior text generation speed and quality. The model boasts 120 billion total parameters with 12 billion active parameters and supports an impressive context length of up to 1 million tokens.

Key Capabilities

• Advanced Agentic Workflows: Designed for collaborative AI agents, enabling sophisticated multi-step reasoning and task execution.
• Long-Context Reasoning: Handles extensive inputs and outputs, supporting up to 1 million tokens for complex document analysis and summarization.
• High-Volume Workloads: Optimized for efficiency in demanding applications such as IT ticket automation and large-scale RAG systems.
• Configurable Reasoning: Users can enable or disable a reasoning trace via the chat template, offering flexibility for different use cases.
• Multilingual Support: Proficient in English, French, German, Italian, Japanese, Spanish, and Chinese.
• Compute Efficiency: Pre-trained using NVFP4 quantization to maximize performance on NVIDIA GPU-accelerated systems, requiring a minimum of 8x H100-80GB GPUs.

Good for

• Developing sophisticated AI agent systems that require robust reasoning and tool-use capabilities.
• Applications demanding long-context understanding and generation, such as legal document review or scientific research.
• Automating high-volume, complex tasks like IT support, customer service, and data processing.
• Building multilingual chatbots and instruction-following systems for global deployment.
• Researchers and developers leveraging NVIDIA's ecosystem for efficient and high-performance LLM deployment.