Google is expanding the Gemma 2 family with three new lightweight, generative AI models: Gemma 2 2B, ShieldGemma and Gemma Scope.

Google is launching three new generative AI models that are said to be safer, smaller and more transparent than most others, the company writes in a blog post. The lightweight AI models Gemma 2 2B, ShieldGemma and Gemma Scope are an addition to the existing Gemma series. Users can now try out the models.

Addition to the Gemma series

The three new models are an addition to the Gemma 2 family, a series of “open” models that were launched earlier this year. The “open” nature should not be confused with open source models. With open source, users have complete creative autonomy without restrictions. In contrast, Google’s Gemini model, where the source code is not made available to developers. With the Gemma series, Google seemed to make a stir among developers.

Now Google is launching three new generative AI models that join the Gemma open models club. In addition to the existing Gemma 2B and Gemma 7B, the following models are added: Gemma 2 2B, ShieldGemma and Gemma Scope.

Small and safe

This trio of new Gemma models are designed for slightly different applications and use cases, but share the same values ​​in terms of security approach. Gemma 2 2B is a lightweight model for generating analytics text. This model can run on a wide range of hardware, including laptops and edge devices.

ShieldGemma is a suite of security classifiers. This model attempts to detect harmful content such as hate speech, harassment or sexually explicit content at the entrance and exit. The model is based on Gemma 2.

Finally, Gemma Scope provides researchers and developers with transparency into the decision-making processes of Gemma 2 models. Google describes the model in its blog as follows: “Gemma Scope works like a powerful microscope, using sparse autoencoders (SAEs) to zoom in on specific points in the model and make its internal workings more interpretable.”

Users can try out these new models now by downloading them via NVIDIA NIM or Google AI Studio.