Google DeepMind has introduced Gemma 3 270M, a 270-million-parameter language model engineered to run entirely on consumer devices such as smartphones, Raspberry Pi boards and web browsers. The open-weight model is the smallest member of the Gemma family and is aimed at developers who need low-latency, privacy-preserving AI without relying on cloud GPUs. Gemma 3 270M combines 170 million embedding parameters with 100 million transformer-block parameters and supports a 32,000-token context window. An instruction-tuned version scores 51.2 percent on the IFEval benchmark, outperforming other sub-billion-parameter models. Google’s internal tests show an INT4-quantised build consuming just 0.75 percent of a Pixel 9 Pro battery over 25 conversations, underscoring the model’s energy efficiency. Google is releasing both pre-trained and instruction-tuned checkpoints alongside Quantisation-Aware-Training variants, with downloads available through Hugging Face, Kaggle, Ollama and LM Studio. The model is covered by the Gemma custom licence, which allows broad commercial use provided safety restrictions are observed. By pairing modest size with strong instruction following, Google positions Gemma 3 270M as a cost-effective foundation for task-specific fine-tuning in areas such as sentiment analysis, entity extraction and offline creative applications.