Google releases ‘open’ AI models after Meta

SAN FRANCISCO (Reuters): Google on Wednesday released new artificial intelligence (AI) models that outside developers potentially can fashion as their own, following a similar move by Meta Platforms and others.

The Alphabet subsidiary said individuals and businesses can build AI software based on its new family of “open models” called Gemma, for free. The company is making key technical data such as what are called model weights publicly available, it said.

The move may attract software engineers to build on Google’s technology and encourage usage of its newly profitable cloud division. The models are “optimized” for Google Cloud, where first-time cloud customers using them get $300 in credits, the company said.

Google stopped short of making Gemma fully “open source,” meaning the company still may have a hand in setting terms of use and ownership. Some experts have said open-source AI was ripe for abuse, while others have championed the approach for widening the set of people who can contribute to and benefit from the technology.

With the announcement, Google did not make its bigger, premier models known as Gemini open, unlike Gemma. It said the Gemma models are sized at two billion or seven billion parameters — or the number of different values that an algorithm takes into account to generate output.

Meta’s Llama 2 models range from seven to 70 billion parameters in size. Google has not disclosed the size of its largest Gemini models. For comparison, OpenAI’s GPT-3 model announced in 2020 had 175 billion parameters.

Chipmaker Nvidia on Wednesday said it has worked with Google to ensure Gemma models run smoothly on its chips. Nvidia also said it will soon make chatbot software, which it is developing to run AI models on Windows PCs, work with Gemma.