Google Corporation has released three new open-source generative artificial intelligence models, positioning them as “safer, more compact, and transparent” compared to others.
- Gemma 2 2B — a model version with 2 billion parameters, featuring security enhancements and offering a “balance of performance and efficiency”;
- ShieldGemma — a set of neural networks based on Gemma 2 for classifying safe content;
- Gemma Scope — a new tool for interpreting models, providing insights into their internal workings.
These are additions to the Gemma 2 family released in June.
Gemma 2 2B is a lightweight model for text creation. It can operate on hardware devices like laptops and is licensed for certain research and commercial applications.
ShieldGemma aims to detect harmful material such as hate speech, harassment, and explicit sexual content. It is used to filter AI prompts and the content generated by AI.
Gemma Scope makes the internal workings of Gemma 2 more interpretable. It “unpacks” complex information into a form that is easier to analyze and understand.
The Gemma AI model line differs from Gemini, as Google does not open the source code for the latter.
Previously, Meta introduced a collection of new open-source AI systems, Llama 3.1. Mark Zuckerberg supported the idea of developing open artificial intelligence.
In July, Google updated the Gemini chatbot and released two neural networks for solving complex mathematical problems.
In February, the corporation introduced the open AI model Gemma.