Gemma 2

Google DeepMind shipped the second Gemma generation in mid 2024 as English text models at three sizes, roughly 2.6, 9 and 27 billion parameters, each offered as a pretrained checkpoint and an instruction-tuned chat version. They draw on the same research as Gemini, were trained on TPUv5p hardware with 2, 8 and 13 trillion tokens respectively, and accept an 8K-token context, which is short by current standards. Weights sit on Hugging Face and Kaggle behind a click-through license, and load in Transformers, Keras and JAX, or community runtimes such as Ollama and llama.cpp. Ollama's default 9B build is about 5.4 GB and fits a mid-range consumer GPU or a recent laptop, while its 27B build is about 16 GB. The Gemma Terms of Use allow commercial deployment but bind you and your downstream users to Google's prohibited-use policy, so this is not an OSI-approved license. Gemma 3 and the Apache-licensed Gemma 4 have since added image input and far longer context, so Gemma 2 now mainly serves existing deployments, fine-tunes built on it, and interpretability work such as the original Gemma Scope sparse autoencoders, which were trained on it.