Google releases pint-size Gemma open AI model (arstechnica.com)

Google has unveiled Gemma 3 270M, a highly compact version of its open-source Gemma 3 AI model, designed to run efficiently on local devices such as smartphones and web browsers. Unlike the trend toward massive, multi-billion parameter models that require cloud-based GPU clusters, this pint-sized Gemma model features just 270 million parameters—dramatically reducing hardware demands while preserving robust generative capabilities. This development marks a significant shift toward lightweight, on-device AI, emphasizing privacy, reduced latency, and energy efficiency. Despite its smaller size, Gemma 3 270M delivers impressive performance for its scale. In tests on a Pixel 9 Pro’s Tensor G4 chip, it handled up to 25 concurrent conversations while consuming less than 1 percent of the battery. Evaluated on the IFEval benchmark, which measures instruction-following ability, it scored 51.2 percent, outperforming other lightweight models with more parameters and approaching the capabilities of much larger models like Llama 3.2. This balance of efficiency and competence makes Gemma 3 270M a notable tool for developers seeking to deploy privacy-conscious, responsive AI applications outside of the cloud, broadening access to powerful AI on everyday devices.