Google introduces two new open LLMs

Google announced the release of its new family of lightweight open-weight models, the Gemma series, not long after unveiling its latest Gemini models. The Gemma models, specifically Gemma 2B and Gemma 7B, were reportedly influenced by the Gemini models and may be used for commercial or research purposes.

Google has not yet released detailed information comparing these models to similar ones from other companies, such as Meta and Mistral. Like Google’s Gemini and earlier PaLM models, Gemma models have a dense decoder-only structure. The tech giant promises to publish performance benchmarks later today on the leaderboard of software hub, Hugging Face.

Developers can access Google’s new Gemma models via readily available Colab and Kaggle notebooks or through integrations with Hugging Face, MaxText, and Nvidia’s NeMo. These models can be run universally after being pre-trained and optimized.

Although Google underscores that the Gemma models are open models, they are not open-source per se. In a recent press briefing, Google’s Janine Banks stressed the importance Google places on open-source, but also pointed out how the company carefully refers to the Gemma models as open models, implying that developers have wide access to adapt and fine-tune them, provided they abide by the specific terms of use.

Tris Warkentin, Product Management Director at Google DeepMind, highlighted that smaller, state-of-the-art models have led to considerable improvements in generation quality over the past year and have made it possible to develop AI applications in innovative ways, including the ability to run inference and tuning on local developer desktops or laptops.

Google is also introducing a new responsible generative AI toolkit for building safer AI applications with Gemma, as well as a debugging tool. It will be interesting to see how the Gemma models fare in real-world settings, considering the competitive landscape.