Nvidia is taking a bold step into the realm of AI model creation, positioning itself as a significant player in the model-making space with the launch of Nemotron 3. The company, which has thrived on its chip sales to firms working on artificial intelligence, released these advanced open models alongside valuable data and tools for developers, signaling a shift in strategy.
As major AI companies like OpenAI and Google increasingly develop their own hardware, this move can be seen as Nvidia’s attempt to safeguard its relevance in the face of competition. Open models play a crucial role in the AI landscape, offering researchers and startups a foundation for experimentation and development. While OpenAI and Google provide open models, their updates have fallen short compared to those from their Chinese counterparts, which have gained significant traction among engineers, as evidence by Hugging Face data.
The Nemotron 3 models, which can be customized and utilized on personal hardware, are described by Nvidia as among the best in the market. CEO Jensen Huang emphasized the importance of open innovation for advancing AI. “With Nemotron, we’re transforming advanced AI into an open platform that gives developers the transparency and efficiency they need to build agentic systems at scale,” he stated.
Nvidia’s transparency stands out in the U.S. tech landscape, as it plans to release the training data for the Nemotron models. This openness is expected to facilitate easier modifications. The company also introduces new tools for model personalization, including a hybrid latent mixture-of-experts architecture tailored for AI agents capable of performing tasks online or on computers. Reinforcement learning libraries are being included too, allowing developers to teach models through simulated rewards.
The Nemotron models are available in three sizes: Nano (30 billion parameters), Super (100 billion), and Ultra (500 billion). The number of parameters indicates a model’s capability and operational complexity, with the largest sizes requiring substantial hardware resources.
Kari Ann Briski, Nvidia’s VP of generative AI software, outlined that open models serve three essential purposes: customization for specific tasks, effective distribution of workloads between different models, and enhancing intelligent responses through simulated reasoning. “We believe open source is the foundation for AI innovation, continuing to accelerate the global economy,” she stated.
Historically, Meta introduced advanced open models named Llama in early 2023, but has since hinted at potentially closing some future releases. This adjustment reflects a broader trend among U.S. companies that have begun to prioritize secrecy in their AI research and strategies.
A report from OpenRouter also revealed that open models comprised about a third of all tokens processed through its systems in 2025. Meanwhile, Chinese companies like DeepSeek and Alibaba have been consistently releasing open models and sharing research insights, creating a more appealing environment for engineers and innovators.
Nvidia’s reliance on its hardware could pose risks as the landscape evolves. With ongoing geopolitical dynamics influencing trade and technology independence, especially in China, there are concerns that the domestic landscape may lean toward local solutions, potentially undermining Nvidia’s market dominance.
