Gemma

Gemma is a family of open-weights language models from Google DeepMind, built on research and technology from the Gemini models. The family spans several generations (Gemma 1, 2, 3, 3n and 4) and sizes, from compact models such as Gemma 3 270M aimed at on-device use up to larger variants for laptops and cloud servers. Gemma 4 is described by Google as built for reasoning and agentic workflows, and includes a 12B unified multimodal model and quantization-aware-trained (QAT) versions for mobile and laptop efficiency.

Beyond the core models, Google lists official variants: EmbeddingGemma, T5Gemma (encoder-decoder), MedGemma (medical text and imaging), TranslateGemma (55 languages), FunctionGemma (function calling), ShieldGemma 2 (content-safety classifier), VaultGemma (differentially private), DiffusionGemma (text diffusion for generation speed), and Gemma Scope 2 for interpretability research. Multimodal input (text and images) is supported in the recent generations, per the GitHub example.

Weights are downloadable, and the google-deepmind/gemma repository provides a JAX library for sampling and fine-tuning (Apache-2.0 for the code). Models can run on CPU, GPU or TPU and are available through Hugging Face and various integrations. Like other LLMs, Gemma can produce inaccurate or offensive output, and Google advises against relying on it for professional advice. Model weights are released under Gemma-specific terms rather than a standard OSI license.