The same technology and research that went into making the Gemini models are also utilized to build Gemma’s open-source AI models.
On Wednesday, February 21, Google unveiled Gemma, a new lightweight open-source family of artificial intelligence (AI) models. Developers and researchers now have access to two Gemma variants: Gemma 2B and Gemma 7B. The tech behemoth claimed to have developed Gemma using the same tools and research as its Gemini AI models. Remarkably, the Gemini 1.5 model was shown just last week. These reduced language models can be used to create AI solutions tailored to specific tasks, and the company permits distribution and responsible commercial use.
Google CEO Sundar Pichai made the announcement in a post on X, formerly known as Twitter. “Gemma is available worldwide starting today in two sizes (2B and 7B), supports a wide range of tools and systems, and runs on a developer laptop, workstation, or @GoogleCloud,” he stated. “Demonstrating strong performance across benchmarks for language understanding and reasoning.” A developer-focused landing page for the AI model has also been created by the company. On this page, users can access code samples and quickstart links for Kaggle Models, quickly deploy AI tools using Vertex AI (Google’s platform for developers to build AI/ML tools), or experiment with the model and connect it to a different domain using Collab (Keras 3.0).
Google stated that both versions of the Gemma AI models are instruction-tuned and pre-trained, highlighting some of its features. Popular data repositories like Hugging Face, MaxText, NVIDIA NeMo, and TensorRT-LLM are integrated with it. The language models can be used with Vertex AI and Google Kubernetes Engine (GKE) on laptops, workstations, or Google Clouds. In order to assist developers in creating safe and ethical AI tools, the tech giant has also published a new ethical Generative AI Toolkit.
According to findings made public by Google, Gemma has surpassed Meta’s Llama-2 language model on several significant benchmarks, including BIG-Bench Hard (BBH), HumanEval, HellaSwag, and Massive Multitask Language Understanding (MMLU). Notably, according to a number of rumors, Meta has already started working on Llama-3.
In the field of artificial intelligence, it has become popular to release smaller language models as open-source tools for academics and developers. There are already open-source solutions for stability, Meta, MosaicML, and even Google with their Flan-T5 models. On the one hand, it fosters the development of an ecosystem by enabling developers and data scientists outside of AI corporations to experiment with the technology and produce original tools. However, it also helps the business because, in most cases, companies provide deployment platforms that require a monthly charge. Additionally, the adoption process by developers frequently draws attention to algorithmic or training data issues that would have gone unnoticed before release, enabling the businesses to enhance their models.
