“Gemma: Open Models Based on Gemini Research and Technology” was released by Google DeepMind on arXiv on March 13, 2024. Gemma is a family of lightweight, openly available models built from the same research and technology that produced Google’s flagship Gemini models. The initial release came in two sizes, roughly 2 billion and 7 billion parameters, with both pretrained base versions and instruction-tuned versions made available.
The paper reports that Gemma outperforms similarly sized open models on 11 of 18 text-based benchmarks, and it places notable emphasis on responsible release, including safety evaluations and guidance for deploying the models. By open-sourcing models derived from its frontier research line, Google entered the open-weight arena that had been dominated by Meta’s Llama and Mistral.
Gemma’s significance is that one of the largest AI labs chose to publish capable open models alongside its closed flagship, broadening the set of high-quality models organizations can run themselves. It was quickly followed by Gemma 2, which pushed the same idea to better performance at practical sizes.