NVIDIA's most powerful open-source weighted AI model: Nemotron 3 Super debuts.

NVIDIA has released Nemotron 3 Super, its most powerful open-source weighted AI model, boasting 120 billion parameters (MoE architecture, with only 12 billion activated during inference). It delivers a 5x increase in throughput and doubles the accuracy compared to its predecessor. The model features an ultra-long context window with 1 million tokens, designed specifically for multi-agent systems to address "context explosion" and "inference tax" problems. Employing a Mamba-Transformer hybrid architecture, it introduces new "potential MoE" technologies and supports multi-token prediction.