Google Launches Gemma: State-of-Art Open Models

Google has announced the launch of Gemma, a new family of lightweight open models, following their recent release of Gemini models.

Today, Google announced the launch of Gemma, a new family of lightweight open-weight models, just one week after releasing the most recent version of its Gemini models. The new Gemma 2B and 7B models, which were “inspired by Gemini,” are now accessible for both academic and commercial use.

Google merely said that these models are “state-of-the-art.” They did not give us a full study comparing their performance to that of other, comparable models, such as Meta and Mistral. We may expect to see the benchmarks later today on Hugging Face’s leaderboard, however the company did mention that these models are dense decoder-only, the same architecture it used for its Gemini models (and older PaLM models).

Gemma comes with pre-built notebooks for Kaggle and Colab, and it also integrates with Hugging Face, MaxText, and Nvidia’s NeMo, so developers can jump right in. These models are portable after pre-training and tuning.

Despite Google’s claims to the contrary, these models are not open-source. Janine Banks, a Google representative, emphasized the company’s dedication to open source during a press briefing before to today’s release. Banks also mentioned that Google uses the term “Gemma models” quite deliberately.

Using the phrase “open models” has been commonplace in the business world, according to Banks. “And it often refers to open weights models, where researchers and developers have wide access to make models better. However, the terms of use, including redistribution and ownership of variants, differ from model to model.” Because of this discrepancy with the conventional definition of open source, we have chosen to call our Gemma models “open models.”

Even though these model sizes are suitable for many use cases, developers can utilize them for inferencing and adjust them as they like, according to Google’s team.

“The generation quality has gone significantly up in the last year,” said Tris Warkentin, head of product management at Google DeepMind. new, state-of-the-art smaller models can do what was once only achievable with very huge models. We are thrilled about the additional possibilities this opens up for AI application development, such as the ability to execute inference and tuning on a local developer desktop or laptop with an RTX GPU or on a single host in GCP with Cloud TPUs.

We will need to observe the Gemma models’ performance in real-world circumstances to determine whether this is also true of Google’s competitors’ open models in this area.

A new responsible generative AI toolkit, including a debugging tool and “guided guidance and essential tools for creating safer AI applications with Gemma,” is also being released by Google alongside the new models.