Gemma 3

Gemma 3 is Google DeepMind's open-weight language model family, built from Gemini research and released in 1B, 4B, 12B, and 27B parameter sizes. The models handle a 128K-token context window, accept image input on the larger variants, support function calling and more than 35 languages, and are designed so that even the 27B model runs on a single GPU or TPU host. The official google-deepmind/gemma repository contains a JAX library, published as the gemma package on PyPI, for inference and fine-tuning: a ChatSampler API handles multi-turn conversation across Gemma generations, and documented workflows cover LoRA as well as full fine-tuning on CPU, GPU, or TPU. Weights are distributed through channels such as Kaggle and Hugging Face and also run in common runtimes outside the JAX ecosystem. Released under the Gemma license, which allows commercial use subject to a use policy, the family targets developers and researchers who want capable models they can host and customize themselves.