Google Launches Gemma: New Open Models for AI Development and Inference

(CTN News) – Google announced the launch of Gemma, a new family of lightweight open-weight models, just a week after introducing the most recent generation of its Gemini models. Starting with Gemma 2B and Gemma 7B, these additional models were “inspired by Gemini” and are now accessible for commercial and research use.

Google did not provide a full paper on how these models compare to similar models from Meta and Mistral, for example, and simply stated that they are “state-of-the-art.”

However, the business clarified that these are dense decoder-only models, which have the same architecture as its Gemini models (and earlier PaLM models), and that the benchmarks will be available later today on Hugging Face’s leaderboard.

Advancements in AI Development: Gemma’s Impact

To get started with Gemma, developers may use ready-made Colab and Kaggle notebooks and connections with Hugging Face, MaxText, and Nvidia’s NeMo. Once pre-trained and tweaked, these models can be used anywhere.

While Google claims these are open models, it’s important to note that they are not open source. Indeed, at a press event before today’s announcement, Google’s Jeanine Banks emphasised the company’s commitment to open source while stating that Google is very intentional about how it refers to the Gemma models.

“[Open models] has become pretty pervasive now in the industry,” Banks said. “And it frequently refers to open weights models, in which developers and researchers have broad access to customise and fine-tune models, but the terms of use — such as redistribution and ownership of developed variants — vary depending on the model’s specific terms of use.

So we saw some differences between what we would typically refer to as open source, and we concluded it made the most sense to refer to our Gemma models as open models.”

That means developers may use the model for inference and fine-tune it as needed, and Google’s team claims that these model sizes are appropriate for a wide range of use cases.

“The generation quality has gone significantly up in the last year,” Google DeepMind product management director Tris Warkentin stated. “Things that would have previously required tremendously big models are now possible with cutting-edge smaller models.

This opens up new ways of designing AI applications, including the ability to execute inference and tuning on your local developer desktop or laptop with your RTX GPU or on a single host in GCP using Cloud TPUs.

That is also true for the open models from Google’s competitors in this field, so we’ll have to see how the Gemma models function in real-world circumstances.

In addition to the new models, Google is launching a new responsible generative AI toolkit, which includes “guidance and essential tools for creating safer AI applications with Gemma,” as well as a debugger.