Nvidia is quietly building its most ambitious AI model yet: Nemotron 4, an open-source model family whose largest variant is expected to exceed one trillion parameters, according to Reuters, citing The Information and multiple employees working on the project.
The move marks a striking escalation for a company that has long profited by selling the silicon behind AI rather than the models themselves. Nvidia already publishes open-weight Nemotron models, but a trillion-parameter flagship would put it in direct competition with the strongest openly released systems on the market — a field now defined by China's DeepSeek, Moonshot AI's Kimi K3 and Meta's Llama line.
Insiders say the largest Nemotron 4 model is expected to have at least one trillion parameters, with a release possible as early as late fall. The strategy is classic Nvidia: strong open models encourage developers to build applications that ultimately consume Nvidia compute. The company does not need Nemotron itself to be its most profitable product — if an open model increases demand for GPUs, networking gear and inference services, Nvidia wins even when someone else ships the final app.
The bet comes as open-weight models reshape the industry's economics. Moonshot's Kimi K3, released in late July, was the largest open-weight model ever at 2.8 trillion parameters, and analysts credited it with wiping billions off the pre-IPO valuations of OpenAI and Anthropic. Nvidia's entry would give open-model developers another heavyweight — one uniquely positioned to optimize for its own hardware stack.
For developers, a genuinely competitive trillion-parameter open model from Nvidia would mean more choice, more control over data and deployment, and downward pressure on inference prices. For Nvidia, it is a hedge that turns its biggest threat — powerful free models eroding proprietary AI margins — into another reason to buy its chips.
