Google today released Gemma, a family of open-weight models built from the same research and technology used to create its Gemini models, marking Google’s entry into the open-weights LLM space alongside models from Meta and Mistral. Available in two sizes, Gemma 2B and Gemma 7B, each with pre-trained and instruction-tuned variants, the models are designed to run on a developer’s laptop or desktop. Accompanying the weights is a Responsible Generative AI Toolkit including safety classifiers, a debugging tool, and best practices guidance.

The company is positioning Gemma as ‘open models’ rather than open source, a distinction Google’s Jeanine Banks made explicit in a press briefing, noting that terms of use for redistribution and ownership vary across the industry. The models support fine-tuning across JAX, PyTorch, and TensorFlow through native Keras 3.0, with ready-to-use Colab and Kaggle notebooks and integrations with Hugging Face, MaxText, NVIDIA NeMo and TensorRT-LLM. Commercial use and distribution are permitted for organizations of any size.

Performance claims are based on unspecified benchmarks; Google says Gemma surpasses ‘significantly larger models’ on key benchmarks, but did not provide a detailed paper at launch. The benchmarks will appear later today on Hugging Face’s leaderboard. The release includes free access via Kaggle, Colab, and $300 in credits for first-time Google Cloud users, and researchers can apply for Google Cloud credits of up to a collective $500,000.