Nvidia is developing a massive open artificial intelligence (AI) model with over 1 trillion parameters. This initiative aims to catch up with Chinese companies in the open AI market while expanding its AI model ecosystem to boost demand for its graphics processing units (GPUs).
According to IT media outlet The Information on August 11, the largest model in Nvidia's next-generation series, 'NemoTron4,' is expected to have at least 1 trillion parameters, targeting performance that can compete with the world's leading open AI models.
NemoTron4 is approximately twice the size of the previously largest model, 'NemoTron3 Ultra,' which was unveiled in June. The number of parameters is a key indicator of the size of an AI model.
However, in terms of size, Chinese companies still lead the market. Moonshot AI's 'Kimi K3' boasts 2.8 trillion parameters, Alibaba's 'Q1 3.8-Max' has 2.4 trillion parameters, and DeepSeek's 'V4-Pro' is reported to have 1.6 trillion parameters.
Having a high number of parameters does not necessarily guarantee superior performance. Nvidia is enhancing its technology to achieve high performance with smaller models while also increasing the size of its models.
On the same day, Nvidia unveiled a lightweight model, 'NemoTron3.5 Lightning,' which has 30 billion parameters. This model utilizes 'distillation' technology to transfer the capabilities learned by larger models to smaller ones, allowing for high performance with fewer computing resources. It can be used for coding, security checks, and AI program applications.
Nvidia has significantly increased its investment in developing its own AI models. It is securing large-scale computing resources by renting AI servers from cloud companies that purchase its AI chips. As of late April, the multi-year cloud service contract amount reached $30 billion (approximately 42 trillion won).
The company is also expanding collaborations with external AI firms. It is sharing AI training data and technology through the 'NemoTron Alliance,' which includes Reflection AI, Cursor, Thinking Machines Lab, and Mistral AI. Naver Cloud joined the alliance in June.
As Nvidia's own AI model performance improves, the likelihood of competing with major clients like OpenAI and Nvidia-invested startups also increases.
Nevertheless, Nvidia anticipates that as the number of open AI models grows, the demand for GPUs necessary for training and operation will also rise. The company aims to expand its influence beyond AI semiconductor sales to include models and software, thereby growing its AI ecosystem centered around its platform.
Details regarding the specifications and release schedule for NemoTron4 have not yet been finalized. The final training is still underway, and it is expected to be completed by late fall at the earliest.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.