Nvidia has unveiled a lightweight open-source artificial intelligence (AI) model it developed in-house. Following its lead role in a letter opposing regulations on open AI, it now appears to be accelerating its own model development. Nvidia is also said to be working on an ultra-large open AI model with 1 trillion parameters.
Nvidia on the 11th (local time) unveiled Nemotron 3.5 Lightning, a mixture-of-experts (MoE) model with 30 billion parameters.
Nemotron 3.5 Lightning can be downloaded and used by corporations without separate license expense or prior approval. The model's weights are also public so users can modify them to suit their purposes.
Nvidia told CNBC that Nemotron 3.5 Lightning was "distilled" from an existing large Nemotron model to reduce size while delivering similar performance. Distillation is a technique that trains a smaller model using the answers from a large AI model as data. Major corporations have used it to develop lightweight models, but the technology has drawn controversy recently amid allegations that Chinese AI firms used it to imitate advanced U.S. AI models.
According to Nvidia, Nemotron 3.5 Lightning generates tokens up to four times faster than comparable open models, and AI agents finish entire tasks about 30% quicker.
Nvidia also released an open tool called Nemo Switchyard that analyzes the difficulty and requirements of a user's prompt and routes the request to an appropriate AI model. Nvidia said using this can cut expense to one-third while maintaining accuracy comparable to using only the highest-performing model.
It is also progressing on developing an ultra-large AI model. Reuters, citing tech outlet The Information, reported that Nvidia is developing a next-generation model, Nemotron 4, with 1 trillion parameters. Training has not yet finished and no release schedule is set, but development could be completed as early as late fall this year.
Kari Briski, Nvidia's vice president for generative AI, said, "We are investing in Nemotron because we believe the world needs state-of-the-art open models that all corporations and nations can access to enhance safety and security and to spur innovation."
Nvidia's active push to expand the open AI ecosystem also aligns with the interests of its core graphics processing unit (GPU) business. Unlike closed models from OpenAI, Anthropic, and Google, open models can be downloaded by users and run or modified on their own servers or data centers. The more open AI is used, the more demand may grow for AI accelerators needed for training and inference.
Regardless of any particular AI model vendor, the mere increase in AI compute works in Nvidia's favor. Nvidia CEO Jensen Huang said in a recent interview with Axios, "Free AI can only be good for hardware, for semiconductors, and for data centers."