Nvidia Launches Its Most Powerful Open-Weight AI Model Nemotron 3 Super: 120B Params, 5x Throughput Boost

Nvidia recently launched its most powerful open-weight AI model to date, Nemotron 3 Super, with 120 billion parameters and a Mixture of Experts (MoE) architecture (only 12 billion active during inference), designed for large-scale complex agentic AI systems. Companies like Perplexity and Palantir have applied it to search, software development, etc.

The model features a 1 million token context window, solving the "context explosion" and "inference tax" issues in multi-agent interactions. Through MoE and Latent MoE technologies, its throughput increases by 5x, accuracy doubles, and inference speed is 4x faster on the Blackwell platform.

Nvidia has open-sourced the model weights and training methods. Developers can access it via Hugging Face, Google Cloud, etc., or deploy it as NIM microservices locally or in the cloud.