NVIDIA Releases Nemotron 3 Super: A 120B Parameter Open-Source Hybrid Mamba-Attention MoE Model Delivering 5x Higher Throughput for Agentic AI
Nemotron 3 Super is an open-source 120-billion parameter model specifically developed to bridge the gap between proprietary and transparent AI through advanced multi-agent reasoning. Leveraging a hybrid MoE architecture (combining Mamba and Transformer layers) and a massive 1-million token context window, the model delivers 7x higher throughput and double the accuracy of its predecessor, making it highly efficient for complex, long-form tasks. Beyond its raw performance, Nemotron 3 Super introduces "Reasoning Budgets," allowing developers to granularly control compute costs by toggling between deep-search analysis and low-latency responses. By fully open-sourcing the training stack—including weights, datasets—NVIDIA is providing a powerful model for enterprise-grade autonomous agents in fields like software engineering......
Full analysis: marktechpost.com/2026/03/11/…
Model on HF: pxllnk.co/ctqnna8
Paper: pxllnk.co/ml2920c
Technical details: pxllnk.co/lbmkemm
@nvidia @NVIDIAAI @NVIDIAAIDev @ctnzr
Mar 12, 2026 · 2:06 AM UTC
1
360



