Google has launched Gemma 3, a collection of lightweight, state-of-the-art open models that build upon the research and technology of the Gemini 2.0 models. The new model, available in a 27 billion parameter variant, introduces multimodality, supporting both vision-language inputs and text outputs, with the ability to handle context windows of up to 128,000 tokens. Notably, Gemma 3 is designed to be 1.6 times faster in fine-tuning while utilizing 60% less VRAM compared to its predecessor. In collaboration with Hugging Face, a free notebook has been released to assist users in fine-tuning Gemma 3 with Generalized Reward Policy Optimization (GRPO), enabling reasoning and understanding of reward functions. Additionally, the model is multilingual and has expanded context capabilities, enhancing its performance in various applications. The model is now hosted on multiple inference APIs, including aiXplain and Inference Net, making it accessible for developers and researchers.