Gemma is a family of open-weight large language models developed by Google/DeepMind, built from the same research and technology as the Gemini models. The family includes multiple sizes and variants (base pre-trained and instruction-tuned "-it" releases) and covers both text-to-text decoder-only models and, in Gemma 3, multimodal text-and-image input models. Gemma models emphasize relatively small model footprints with strong performance on generation, reasoning, question answering and summarization, and include large-context capabilities (Gemma 3 supports very large context windows). Models and artifacts are published with model cards, technical reports, and tooling (Hugging Face pages, Vertex Model Garden, GitHub repositories), and there are native runtime integrations such as a Unity plugin and C++/DLL bindings for local inference and embedding in applications. Gemma models were trained on modern TPU hardware and have documented safety testing and evaluation materials.