Google today announced the start of Gemma, a new family of lightweight open-weight models. This comes just one week after the latest version of its Gemini models was released. These new models, starting with Gemma 2B and Gemma 7B, were “inspired by Gemini” and can be used for business or study.
Google didn’t give us a full study on how these models compare to other models like those from Meta and Mistral; they only said that they are “state-of-the-art.” This is the same design that the company used for its Gemini models and its earlier PaLM models, the company did say that these are dense decoder-only models. The benchmarks will be posted on Hugging Face’s leaderboard later today.
Developers can use ready-made Colab and Kaggle notebooks to get started with Gemma. It also works with Hugging Face, MaxText, and Nvidia’s NeMo. They can run anywhere after being trained and tuned once.
It’s important to note that these are not open source models, even though Google says they are. In fact, Google’s Jeanine Banks stressed the company’s commitment to open source in a press briefing before today’s release. She also said that Google is very careful about how it talks about the Gemma models.
“[Open models] are pretty common now in the business world,” Banks said. “And it usually means open weights models, where developers and researchers can change and improve models without limits. However, the model’s own terms of use determine the rules for things like redistribution and who owns the new versions that are made.” We think that calling our Gemma models “open models” makes the most sense because they are different from what we usually mean by “open source.”
That means devs can use the model to draw conclusions and make changes as they see fit. Google’s team says that these model sizes are good for many situations.
Tris Warkentin, head of product management at Google DeepMind, said, “The generation quality has gone up a lot in the last year.” “Things that used to only be possible with very large models are now possible with cutting edge smaller models.” We’re really excited about this because it opens up whole new ways to make AI apps. For example, you can now run inference and tuning on your local developer PC or laptop with your RTX GPU, or you can use Cloud TPUs on a single host in GCP.
That’s also true of open models from Google’s rivals in this field, so we’ll have to wait and see how the Gemma models work in real life.
Also Read: We Put Google’s Gemini Robot to the Test. This is How It Did
As well as the new models, Google is also releasing a debugging tool and a new responsible generative AI toolkit that will give “guidance and essential tools for creating safer AI applications with Gemma.”
What do you say about this story? Visit Parhlo World For more.
