Google Adds Two New Models to the Gemma Family

Google today released two new members of its Gemma family of lightweight, open AI models aimed at code completion and research experiments, respectively.

“Gemma is our family of lightweight, state-of-the-art open models built from the same research and technology used to create the Gemini models,” Google director Tris Warkentin and senior product manager Jane Fine write in the announcement post. “Today, we’re excited to announce our first round of additions to the Gemma family, expanding the possibilities for ML developers to innovate responsibly.”

Google announced Gemma in February in two weights and in pre-trained and instruction-tuned variants. Unlike Gemini, Gemma models are designed to be run locally on laptops, PCs, and workstations in addition to the cloud, and they’re optimized for multiple AI platforms.

Today’s additions include CodeGemma, which provides code completion, code generation, and chat for developers and businesses, and RecurrentGemma, a memory efficient model that offers higher throughput for deep learning research. CodeGemma utilizes the Gemma 7B model weight and is available in the same two variants as the original Gemma models, while RecurrentGemma is based on Gemma 2B.

Google also released Gemma 1.1 today with performance improvements, bug fixes, and more flexible licensing terms.

You can learn more about Gemma at the Google AI for Developers website.