(Credit: NurPhoto / Contributor / Getty Images)
Google has launched two new open AI models, Gemma 2B and Gemma 7B, with the goal of supporting AI use in third-party applications.
The Gemma models vary in size, meaning the simpler Gemma 2B may be better for less intensive AI applications, while Google data suggests that Gemma 7B could outperform other AI models like Mistral or Meta's LLAMA-2.
The Gemma models were "built from the same research and technology used to create the Gemini models," Google said in its announcement. Gemini is Google's rebranded Bard AI tool, which is able to generate written content and answer questions.
Third-party developers can download Gemma's files, tinker with them, and add them to their products as long as they are used responsibly, Google says. While Gemma isn't fully "open source," its rules seem to fall somewhere between a closed-off product and an open-source one.
"Its license gives users full creative autonomy," Google said in a blog post Wednesday. "This is a powerful guarantee of technology access for developers and end users."
Google claims that open-source concepts "can't always be directly applied to AI systems," suggesting a possible reason why Gemma was released with the more restrictive "open" categorization instead.
While Google didn't specify exactly what responsible Gemma AI use looks like, the company has established its own AI principles that include prohibiting the creation of illegal content with AI and barring its AI products from being trained on personal information, to name a few examples.
To access Gemma, developers and prospective users have to fill out an access request form and agree to a consent form.
Nvidia says it also worked "closely" with Google to fine-tune the Gemma models and that Nvidia's large language model (LLM) TensorRT was used in the development process. Ahead of its Wednesday earnings call, Nvidia announced in a blog post that it will add Gemma to its recently launched Chat With RTX feature soon.