vidia continues to step up its open-source game.

On Wednesday, the company debuted Nemotron 3 Super, a 120 billion parameter open model designed to run complex agentic AI systems. The model features advanced reasoning, a mixture-of-experts architecture, a 1 million token context window and is built to be more efficient, more accurate and faster than its predecessor.

Nemotron 3 Super also outranked several models from OpenAI, Amazon and Google on the Artificial Analysis benchmark, and can be 2.2 times faster than GPT-OSS in reasoning workloads, Bryan Catanzaro VP of applied deep learning at Nvidia, told The Deep View. It’s the second release in the Nemotron family, following the release of Nemotron Nano in December, with a “four times bigger” Ultra model being released soon, he said.

Nvidia’s latest model comes as open models catch an increasing amount of attention, especially Chinese models like DeepSeek and Qwen.

• The difference, however, is that Nvidia 3 Super isn’t just open weights, said Catanzaro: it’s open “data and recipes.”
• Along with releasing the model itself, the company released the entire methodology for training it, including the pre- and post-training data sets, the training environments and the evaluation recipes.
• “The reason why we're doing this is that we're trying to help the ecosystem,” said Catanzaro. “We work with every AI company, small and large, old and young. We know that it's in our interest to help the ecosystem grow, because it creates opportunity for us.”

And Nemotron 3 Super is proving itself in practice, too. CrowdStrike, which had early access to the model, found that it performed three times more accurately than the previous model it was using in production and performed exceptionally well on internal benchmarks for threat hunting, Sven Krasser, chief scientist at CrowdStrike, told The Deep View.

“We're very excited that something with these capabilities is out there in the open to use,” Krasser said.

These models may just be the beginning of Nvidia’s contribution to the open source ecosystem. According to Wired, Nvidia intends to spend $26 billion on building open models over the next five years, further entrenching the company into the open model ecosystem.

Our Deeper View

What goes around, comes around. As Catanzaro said, feeding the ecosystem is a long-haul investment for Nvidia. As the hardware provider of choice for AI companies, creating one of the most open and flexible models in the ecosystem aligns better with its long-term business goals than it does for Meta, OpenAI or other open model providers in the US. The more Nvidia supports the open-source AI ecosystem, the more companies will customize models for their use cases and boost their AI use, in turn creating more Nvidia customers. Nvidia’s bet also feeds the open source market in North America, one that’s particularly lacking as open models from Chinese firms have recently taken the lead.