Google’s DeepMind research arm on 14 August unveiled Gemma 3 270M, an open-weight language model that packs 270 million parameters into a download of roughly 240 MB. The model is the smallest member of the Gemma 3 family and is aimed at developers who need fast, low-cost inference on laptops, edge devices and smartphones. Built on the Gemma 3 architecture, the system allocates 170 million parameters to embeddings and 100 million to transformer blocks, backed by a 256 k-token vocabulary to handle rare terms. An instruction-tuned variant scored about 51 percent on the IFEval benchmark, outperforming several models with similar or larger footprints, according to figures released by Google. Energy consumption is a principal selling point: internal tests show an INT4-quantised version consuming just 0.75 percent of a Pixel 9 Pro’s battery across 25 conversational sessions. The company says Gemma 3 270M supports rapid fine-tuning in minutes and can run fully offline in browsers, Raspberry Pi boards and other low-power hardware. Google is releasing both pre-trained and instruction-tuned checkpoints, plus Quantization-Aware Training versions, under its Gemma licence that permits commercial use subject to safety restrictions. Documentation and deployment recipes for tools such as Hugging Face, Ollama and JAX are provided to accelerate adoption by enterprises and independent developers.