Why Nvidia Is Trying To Develop The World’s Best Open-Source AI Models

Nvidia is doubling down on its open-source efforts, pouring investment into an ambitious new in-house AI model that it hopes will drive demand for its hardware—but that could also compete with customers and partners.

Nvidia is aiming for the new model, the biggest of a family that it is calling Nemotron 4, to be on par in its performance with the best open-source AI models in the world, according to several people who work on Nemotron. The research paper detailing Nvidia’s last big model had 570 authors, and employees said Nemotron 4 would involve even more people. “Everyone just wants to be involved at this point,” one former employee said.

That work builds on a spate of recent open-source releases from Nvidia. On Tuesday, Nvidia released Nemotron 3.5 Lightning, a smaller model that it said is designed to run AI agents as fast and efficiently as possible. The company also released free model routing software to help companies more easily make their own model routers—tools that direct specific AI tasks to the best, cheapest models for those tasks.

Multiple employees working on Nemotron said they expect the biggest Nemotron 4 model to have at least a trillion parameters—the numbers in a model that are adjusted as it learns. That’s approximately double Nvidia’s current largest model, Nemotron 3 Ultra, released in June. A trillion parameters would still be far smaller than leading Chinese open-source models, but Nvidia also heavily emphasizes compression techniques that can enable smaller models to outperform.

In a sign of Nvidia’s open-source commitment, it has sharply increased spending on the compute used to train its own models. Nvidia gets compute by renting AI servers back from some of the cloud operators that buy its chips. As of April, it had increased the value of such multiyear cloud-service commitments to $28 billion through the start of 2031, approximately triple the amount it reported a year earlier. Some of that cloud usage also goes to other research and development efforts.

The accelerated Nemotron push puts Nvidia in a delicate position of, in effect, competing with both a collection of open-source startups that it has invested in and with the big AI labs that have been Nvidia’s biggest customers. OpenAI in particular has been one of the biggest drivers of demand for chips from Nvidia, which has invested $30 billion in the AI lab.

Though the top Nemotron 4 model isn’t expected to be as powerful as the frontier models from OpenAI and Anthropic, such open-source models are gaining traction among companies as a lower-cost alternative for some work.

Nvidia also has invested sizable amounts in U.S. open-source companies that hope to get companies to use their own models, including Reflection AI and Thinking Machines Lab.

Nvidia is ramping up its open-source efforts because a diversity of open models will spur more demand for its chips, according to conversations with more than a dozen current and former employees, Nvidia partners and Nemotron users. Much of Nvidia’s current demand comes from a handful of frontier labs and cloud operators such as OpenAI, Microsoft and SpaceX, some of which are developing their own AI chips.

Nvidia can benefit, the thinking goes, if a broader range of companies, from startups to large legacy businesses, can use and build AI via high-quality, relatively low-cost models from Nvidia optimized for its hardware—or if its efforts spur faster development of other open-source models.

“Nvidia wins no matter which of these companies makes great open source,” said Anastasios Angelopoulos, chief executive of Arena, a company that evaluates and benchmarks AI models.

Stoking More Open-Source Competition

Nvidia believes its model development efforts will stoke competition and lead to more open model development and therefore more GPU demand, an Nvidia executive said. The company has included several U.S. open-source developers in a network of firms contributing training data and ideas as part of what it calls the Nemotron Coalition to develop new models.

“Nvidia is investing in Nemotron because we believe every company and every country needs accessible frontier open models to strengthen safety and security, accelerate innovation, and provide a foundation they can rely on from one generation to the next,” Kari Briski, Nvidia’s vice president of generative AI, said in an email.

The timing for Nemotron 4 is unclear. Several employees said Nvidia has made some decisions about pretraining data and architecture for the big Nemotron 4 model, but hasn’t set the exact specifications or release date. It also has yet to do the final training run, which could take months. Two employees said the model could be ready late this fall, while others said it could be later.

Nvidia has long developed its own AI models to better inform how it designs its hardware. Early last year, it realized American open-source models had fallen far behind the leading Chinese counterparts, and wanted to help bridge the gap given the U.S. national interest and because it would benefit Nvidia, two employees said.

Bryan Catanzaro, Nvidia’s vice president of applied deep learning research, said on a podcast in January that investing in Nemotron is “essential for the future of our company.”

Open-source AI has become a geopolitical lightning rod as increasingly powerful Chinese models fueled debate in the U.S. over potentially regulatory action. Nvidia CEO Jensen Huang has loudly supported open-source, saying in an email last month that open models “promote safety, cybersecurity, scientific advances, and national security.”

The Nvidia models have gained adoption with some companies, notably including Palantir, which in June announced a collaboration with the chip company to use Nemotron models for U.S. government clients.

But the models’ performance has ranked well behind the industry vanguard. Nemotron 3 Ultra, Nvidia’s biggest model to date, ranks second among U.S. open models on multiple performance benchmarks from Arena and Artificial Analysis on agentic performance and intelligence, respectively, behind Thinking Machines’ just-released Inkling. That is far behind the top Chinese open model and outside of the top 40 models overall.

There is a limit to how much even Nvidia can invest in model development given the other demands on its finances. Nvidia’s cloud services commitment amounts to $7 billion for the fiscal year that will end in January 2028. While that is dwarfed by spending from OpenAI and Anthropic, the amount is still more than most leading open-source labs, both American and Chinese, have raised in their lifetimes. One employee said Nvidia is trying to get more compute for Nemotron.

The Neomotron Coalition companies—which include Reflection, Cursor, Thinking Machines and Mistral—are also contributing training data, evaluation help and model design ideas to Nemotron 4 even as they work on their own open-source projects. Startup Prime Intellect, which helps customers train, fine-tune and run their own open AI models, contributed 300,000 simulated environments to help with model training, said someone directly involved in the collaboration. The AI coding-tool startup Cognition, a member of the coalition whose participation hasn’t yet been announced, has discussed giving Nvidia coding data for training, according to someone with direct knowledge of the talks.

The partners have various incentives to participate. For those working on their own models, Nvidia’s push could expand interest in open source and thereby help build the market. Others are hoping to influence the development of a model they can use without having to pay for the compute that trains it.

Vincent Weisser, CEO of Prime Intellect, sees the coalition as a way to band together against the existence of “one ‘God’ model,” rather than a game of who can make the best open model.