Google Gemma

Google Shares Open-Source Gemma Generative AI Models as Lightweight Gemini Alternative

Google has expanded its generative AI model portfolio beyond the Gemini series with a new line of open-source large language models (LLMs) called Gemma. The Gemma 2B and 7B models (referring to their parameter size) are aimed at developers interested in Gemini who prefer a more lightweight yet accessible option.

Gemma AI

Gemma is Latin for precious stone and the origin of the English word gem. Google claims Gemma refers to how valuable the models are. The similarity in sounds to Gemini is likely a deliberate choice by the company, despite the Greek reference to mythical twins being otherwise unrelated. Both sizes of Gemma models have pre-trained and instruction-tuned variants. The pre-trained models can be used immediately for most tasks, while the instruction-tuned ones have been optimized for what Google calls a safer and more reliable performance. The models can utilize common frameworks like PyTorch, TensorFlow, and JAX for tasks like fine-tuning a custom dataset. Google boasts that Gemma achieves best-in-class performance compared to other publicly available models of equivalent sizes and shared some test results confirming that claim.

“Gemma models share technical and infrastructure components with Gemini, our largest and most capable AI model widely available today. This enables Gemma 2B and 7B to achieve best-in-class performance for their sizes compared to other open models. And Gemma models are capable of running directly on a developer laptop or desktop computer,” Jeanine Banks, Google vice president and general manager for Developer X and head of developer relations, explained in a blog post. “Notably, Gemma surpasses significantly larger models on key benchmarks while adhering to our rigorous standards for safe and responsible outputs. See the technical report for details on performance, dataset composition, and modeling methodologies.”

Flexible Safety

Gemma models are also flexible enough to run on anything from a laptop to a high-powered cloud computing system. Google is working with Nvidia to optimize Gemma on Nvidia’s GPUs. For cloud deployment, the models are tuned to leverage Google Cloud’s Vertex AI and infrastructure. The cross-platform design enables Gemma models to scale from small projects on a single computer to enterprise-level implementations leveraging clusters of GPUs or TPUs on Google Cloud.

Google also ran extensive tests to get Gemma to meet the company’s safety and responsibility standards, highlighting the use of techniques like data filtering and reinforcement learning from human feedback to get Gemma to fit its rules for responsible behavior. The company is eager to encourage the same approach from future Gemma users, leading to another new product released alongside the models. Google’s new Responsible Generative AI Toolkit packages best practice guidelines and programs for developers to help debug models, set up safety guardrails, and investigate potential issues with Gemma-derived creations.

Google painted a vision for Gemma where developers in healthcare, education, and other public-minded spaces can create new generative AI applications faster and with lower resource demands. The smaller size of Gemma compared to Gemini makes the models less powerful, though easier to use in some contexts. By comparison, the new Gemini Ultra model has 1.56 trillion parameters, and the new Gemini 1.5 Pro model has a context window of up to a million tokens in some tests. To spur adoption, Google is offering free access to Gemma through services like Kaggle and Colab notebooks. The company said researchers can also apply for Google Cloud credits up to $500,000 to accelerate projects built on Gemma.

“Gemma is built for the open community of developers and researchers powering AI innovation,” Banks wrote. “As we continue to expand the Gemma model family, we look forward to introducing new variants for diverse applications.”


Google Releases Gemini Generative AI Model Designed to Match and Exceed OpenAI’s GPT-4 and ChatGPT

Google Unveils Gemini 1.5 Pro LLM With Staggering 1M-Token Context Window

Google Bard (and Duet and Assistant Mobile App) No More – Gemini Now The Star of Generative AI Show