Nvidia invests in OpenAI and publishes competitor model
NVLM 1.0 is a family of language models that, according to Nvidia, can compete with GPT-4o. The models are multimodal and open source.
(Image: Konstantin Savusia/Shutterstock.com)
Nvidia is releasing its own family of Large Language Models (LLMs). NVLM 1.0 is led by a 72B parameter model with the name NVLM-D 72B. This should be able to compete with OpenAI's GPT-4o. Nvidia is also investing directly in OpenAI, thereby fueling its own competition.
NVLM is made available by Nvidia Open Source – both weights and the training code are made available via Megatron-Core. This is Nvidia's open-source library.
In the benchmarks, the results of which Nvidia has published in a blog post, NVLM-D 72B consistently performs similarly well to the current versions of Meta's open-source model Llama, as well as the closed models OpenAI GPT-4o, Anthropic Claude and Google Gemini. "The results show that NVLM 1.0 achieves performance comparable to leading models on both vision and text-only tasks," writes Nvidia. In the math tasks, the model even performed better than GPT-4o in some cases. Even after multimodal training, NVLM-D 72B maintained or even improved its performance on text-only tasks. It is a common problem that models perform worse on text-only tasks after multimodal training.
The Nvidia model showed a good ability to follow instructions. It can also generate a very "high-quality, detailed description of the given image", writes Nvidia.
Competition on the open-source AI market
Just a few days ago, it was announced that Nvidia is investing around 100 million US dollars directly in OpenAI. OpenAI will receive a total of 6.6 billion US dollars. The largest investor is a New York venture capitalist called Thrive Capital. Of course, Microsoft is also providing fresh money – around one billion US dollars. In return, the investors are demanding that the company's non-profit status be replaced by new profit-oriented structures. Nevertheless, OpenAI's finances are in a difficult position, and it is clear that income does not cover expenditure.
Alongside Nvidia, Meta in particular also offers its own AI models as open source. Mark Zuckerberg himself repeatedly emphasizes that he sees open-source as the only viable way for AI to develop and an entire ecosystem to emerge. Of course, he wants to be involved in this. With Nvidia, a competitor has now entered this AI open-source market. Zuckerberg also says that Meta can afford to make its own models freely available because they have other sources of income. This also applies to Nvidia, which earns a lot of money with its AI chips. And who will ultimately also benefit if OpenAI needs more of these chips.
(emw)
