Google unveils ultra-small and efficient open source AI model Gemma 3 270M that can run on smartphones

Today, Google's DeepMind AI research team introduced the Gemma 3 270M, an innovative open-source AI model that boasts a modest 270 million parameters. This size is significantly smaller compared to many advanced large language models (LLMs) that often feature over 70 billion parameters. While larger models typically promise enhanced capabilities, Google's new model is designed with efficiency in mind, allowing it to operate directly on smartphones without requiring an internet connection. Internal tests conducted on a Pixel 9 Pro System on Chip (SoC) demonstrated that Gemma 3 270M can tackle complex, domain-specific tasks and can be fine-tuned in just minutes to meet the needs of developers, whether they are part of a large enterprise or an indie startup. Omar Sanseviero, an AI Developer Relations Engineer at Google DeepMind, emphasized that this model is versatile enough to run in various environments, including web browsers and Raspberry Pi devices, highlighting its capability to function on lightweight hardware. One of Gemma 3 270M's standout features is its energy efficiency. In tests utilizing the INT4-quantized model, the device consumed only 0.75% of battery power after 25 conversations. This makes it an appealing option for on-device AI applications, particularly where privacy is a concern and offline functionality is crucial. The model is released in both pretrained and instruction-tuned versions, offering immediate utility for various instruction-following tasks. With a strong performance in the IFEval benchmark, which assesses a model's instruction-following capabilities, Gemma 3 270M achieved a score of 51.2%. This places it above other small models such as Smol LM2 135M Instruct, and it competes closely with larger models, showcasing Google's commitment to optimizing smaller architectures for specific applications. Additionally, the model’s architecture supports rapid deployment on devices with limited resources, making it suitable for various tasks from sentiment analysis to creative writing. Developers have access to comprehensive documentation, fine-tuning recipes, and deployment guides to facilitate smooth transitions from experimentation to production. In a demonstration video, Google showcased a Bedtime Story Generator app utilizing Gemma 3 270M, which operates entirely offline within a web browser. This app illustrates the model's ability to generate engaging narratives based on user-selected parameters, proving its potential in creative applications. Gemma 3 270M is governed by specific terms of use that permit reproduction, modification, and distribution, provided that developers adhere to Google's guidelines. This flexibility enables businesses to integrate the model into their products without additional licensing fees while ensuring compliance with relevant laws. With the Gemmaverse exceeding 200 million downloads, Google positions the Gemma 3 270M as a critical tool for developing efficient, privacy-conscious AI solutions across various platforms, marking an exciting advancement in the field of mobile AI technology.