NVIDIA's Nemotron 3 Super Rewrites the Rules of Open-Source Agentic AI

News Summary

On Wednesday, March 11, 2026 (Eastern Time), NVIDIA officially unveiled Nemotron 3 Super, a powerful new open-source large language model specifically engineered for autonomous, multi-agent AI applications. The release marks a significant milestone in NVIDIA's broader strategy to lead the open-source AI landscape in the United States, backed by a reported $26 billion investment in open-model AI development.

A 120-Billion Parameter Powerhouse

Nemotron 3 Super is a 120-billion-parameter model that operates with only 12 billion active parameters during inference, thanks to its hybrid Mixture-of-Experts (MoE) architecture. The design combines Mamba-2 state-space layers with Transformer attention layers, enabling the model to handle long-context tasks with remarkable memory and computational efficiency. It delivers over 5x the throughput of its predecessor, Nemotron Super, and claims 2x higher accuracy compared to the previous generation.

Solving the Two Core Problems of Agentic AI

NVIDIA built Nemotron 3 Super to directly address two persistent bottlenecks in multi-agent AI systems. The first is "context explosion" — the exponential token growth that occurs when agents re-send tool outputs, reasoning steps, and conversation histories with every turn. To counter this, Nemotron 3 Super features a native 1-million-token context window, allowing agents to maintain full workflow awareness over extended tasks without losing alignment with their original goals.

The second is the "thinking tax" — the prohibitive cost of deploying large reasoning models for every subtask in complex workflows. The model's hybrid MoE architecture and three core innovations tackle this head-on: Latent MoE (which activates four times as many expert specialists at the same compute cost), Multi-Token Prediction (which predicts multiple tokens per forward pass, cutting generation time significantly), and native NVFP4 precision training (enabling up to 4x faster inference on NVIDIA Blackwell GPUs compared to FP8 on Hopper).

Fully Open and Enterprise-Ready

Unlike many high-performance models, Nemotron 3 Super is released with fully open weights, datasets, and training recipes under a permissive license. NVIDIA is publishing over 10 trillion curated pre-training tokens, 40 million post-training samples, and reinforcement learning environments across 21 configurations. Developers can access the model on Hugging Face, build.nvidia.com, OpenRouter, and Perplexity, and deploy it via NVIDIA NIM microservices.

Enterprise access is available through Google Cloud Vertex AI and Oracle Cloud Infrastructure, with Amazon Bedrock and Microsoft Azure deployments coming soon.

Industry Adoption and Benchmark Performance

Leading AI-native companies have already integrated the model. Perplexity offers Nemotron 3 Super as one of 20 orchestrated models in its Computer product. Software development agent companies CodeRabbit, Factory, and Greptile are deploying it to improve accuracy while reducing compute costs. Enterprise heavyweights including Palantir, Cadence, Siemens, and Amdocs have also committed to integration.

On the competitive benchmark front, Nemotron 3 Super has claimed the top position on Artificial Analysis for efficiency and openness. NVIDIA's AI-Q research agent, powered by Nemotron 3 Super, reached No. 1 on both DeepResearch Bench and DeepResearch Bench II. The model scored 85.6% on PinchBench, outperforming Anthropic's Claude Opus 4.5, Kimi 2.5, and GPT-OSS 120B. In throughput comparisons, it is 2.2x faster than OpenAI's GPT-OSS 120B and 7.5x faster than Alibaba's Qwen3.5-122B.

The Bigger Picture: Open AI as National Strategy

NVIDIA CEO Jensen Huang emphasized the strategic importance of open innovation, stating that the company aims to transform advanced AI into an open platform that gives developers the transparency and efficiency needed to build agentic systems at scale. Analysts note that the Nemotron 3 release is also a geopolitical play — as China's Qwen and Kimi models have dominated the open-source space, Nemotron positions the United States with a competitive, fully transparent alternative.

Nemotron 3 Super sits in the middle of the Nemotron 3 family, sitting above the previously released 30-billion-parameter Nemotron 3 Nano (launched December 2025) and below the anticipated 500-billion-parameter Nemotron 3 Ultra, whose release date has not yet been announced.