Nvidia just rolled out Nemotron 3 Super, an open-source AI model which has been built for creating advanced autonomous AI agents. It’s the latest addition to the Nemotron 3 lineup, and it’s made to handle tricky workflows where lots of AI agents need to work together and bounce ideas off each other.
The company has said that this new model is capable of tackling tough reasoning problems, and it will still keep its answers accurate. They are making the weights open, so developers and organisations can grab the model, run it locally, and really dig in. You will find it on Hugging Face, Nvidia’s own site, OpenRouter, and even Perplexity’s AI platform.
So, what’s the big deal with agentic AI systems?
Basically, you have several AI agents teaming up to do things like research, planning, coding, or automating tasks. These setups usually demand a ton of computing power, because you need every step in the process to remember all the details.
Nemotron 3 Super tries to tackle that head-on. With large context windows and advanced reasoning capabilities, it enables AI agents to track long, complex workflows and make decisions across multiple steps without losing their place.
Perplexity’s already using Nemotron 3 Super in its new AI-powered computer platform, which shows the model is not just hype; it is out there doing real work.
Nemotron 3 Super
Digging into the tech, Nemotron 3 Super uses a hybrid mixture-of-experts (MoE) architecture. That’s a fancy way of saying the model only “wakes up” the parts it needs at any given time, which saves on computing power. The specs are impressive: 120 billion parameters overall, with 12 billion active at once, and a context window that spans a whopping 1 million tokens. That’s a significant memory requirement for AI agents managing long, complex tasks.
Latent MoE
There is also a new twist called Latent MoE. With this, the model can tap into four expert modules for the price of one when generating tokens. This means it is both smart and efficient – and it further cuts down the cost of heavy thinking without compromising the performance. These kinds of upgrades help make advanced AI practical for big companies and real-world applications.
Training-wise, Nemotron 3 Super learned from synthetic data created by top-tier reasoning models. The training process chewed through more than 10 trillion tokens across all the datasets.
Nvidia’s also put out detailed docs covering how they trained the model, the reinforcement learning setups they used, and how they tested everything. By sharing all this, they’re hoping developers and researchers can use Nemotron 3 Super as a foundation for building even better AI systems.