NVIDIA has once again made a stunning impact in the field of AI large models. On March 12, NVIDIA officially released a new generation of open-source large model specifically designed for AI agents - Nemotron 3 Super. With its extremely high reasoning efficiency and excellent task success rate, this model quickly became the focus of the open-source community.

Architecture Innovation: Reasoning Speed Increased by 300%

Nemotron 3 Super adopts an innovative Mamba-MoE hybrid architecture, with a total of 120 billion parameters and only 12 billion activated parameters. This design allows it to maintain powerful performance while increasing reasoning speed by three times and throughput by five times. In addition, the model supports an ultra-long context of up to 1 million, effectively solving common issues such as "goal drift" and "context explosion" in multi-agent collaboration.

Performance Showdown: The "Performance Ceiling" in the Open-Source World

In multiple authoritative evaluations, Nemotron 3 Super performed outstandingly. It not only topped the efficiency and openness rankings of Artificial Analysis, but also drove NVIDIA's self-developed AI-Q agent to first place on both lists of DeepResearch Bench. Notably, the model achieved a success rate of 85.6% in the popular agent task OpenClaw, with performance approaching closed-source large models such as Claude Opus 4.6 and GPT-5.4.

Compatibility with the "Blackwell" Platform: Support for NVFP4 Training

To fully leverage the advantages of its self-developed hardware, Nemotron 3 Super supports not only BF16 and FP8 formats, but also specifically supports NVFP4 training for NVIDIA's latest Blackwell platform and subsequent architectures. This feature will further reduce the training cost of large models and improve computing power utilization.

Ecosystem Integration: Major Companies Have Integrated It

Currently, Nemotron 3 Super has been integrated by several technology giants including Perplexity, Palantir, Siemens, and Dell, and is also available on mainstream cloud platforms such as AWS, Azure, and Google Cloud. As an open-source and free model, it provides developers with a low-cost and high-performance alternative, greatly impacting the current market landscape dominated by closed-source large models.