What Happened
NVIDIA released Nemotron 3 Super on March 11, 2026. Although the Nemotron 3 family had been announced earlier, the March post explicitly marks this model’s availability. NVIDIA provided downloadable weights alongside training resources and deployment integrations for agent systems.
Why It Matters
The release expanded access beyond a finished model checkpoint. Publishing datasets, reinforcement-learning environments and evaluation recipes gave researchers more material for understanding and reproducing model development. It also made hybrid sequence architectures a practical option for long-running agents.
Technical Details
Nemotron 3 Super has 120 billion total parameters, with 12 billion activated, and supports a one-million-token context window. Its architecture combines Mamba components, attention and mixture-of-experts layers. LatentMoE improves expert utilization, while multi-token prediction supports faster decoding. NVIDIA also optimized the model for NVFP4 operation on Blackwell hardware. Vendor-reported speedups depend on the specified hardware and workload.