From the archive
Google releases Gemma 2B and 7B model weights
Gemma brings downloadable pretrained and instruction-tuned models to local machines and cloud deployments.
Google DeepMind and other Google teams released the first Gemma models on February 21, 2024. The downloadable models came in 2-billion and 7-billion parameter sizes, each with pretrained and instruction-tuned variants.
Local and cloud deployment
The release included integrations with Hugging Face , Colab, Kaggle, and common model frameworks. Gemma was designed for deployment on local computers as well as Google Cloud , including Vertex AI and Google Kubernetes Engine.
Unlike the hosted Gemini API, Gemma supplied model weights for running and adapting the models in a chosen environment. Its terms permitted commercial use and distribution, subject to the Gemma terms of use.
Tools around the weights
Google also released a Responsible Generative AI Toolkit with guidance and tools for testing model behavior. The launch added a smaller downloadable model family to Google's hosted AI offerings, with infrastructure and model operation left to the deployment.