Gemma 4 hits the edge for privacy

A language model is a prediction engine for words, and the smallest versions now fit on consumer hardware instead of rented cloud servers. Google said last week that Gemma 4 includes E2B and E4B variants built for phones, laptops, and other edge devices. (blog.google) Google announced four Gemma 4 sizes on April 2: E2B, E4B, 26B A4B, and 31B. Its model card says the E2B and E4B versions are meant for “high-end phones,” laptops, and similar hardware, while the larger models target personal computers and servers. (ai.google.dev) Running a model locally means the prompt and output stay on the device unless an app sends them elsewhere. Google’s Android developers blog said Gemma 4 can run in Android Studio with “the model and inference contained entirely on your local machine.” (developer.android.com) Google is pitching that setup as an alternative to cloud-only assistants that process requests on third-party infrastructure. Its developer blog said Gemma 4 is designed for “on-device agentic workflows” across mobile, desktop, and internet-connected hardware, with support for more than 140 languages. (developers.googleblog.com) The tradeoff is scale. Google’s launch post says local inference is best for offline use, but it also says Google Cloud “removes all compute ceilings,” a reminder that smaller on-device models still give up some raw capacity compared with larger hosted systems. (blog.google) Google released Gemma 4 under the Apache 2.0 license, which allows broad commercial use with attribution and license notice requirements. The company says developers can run the models on their own hardware, mobile devices, or hosted services, then fine-tune them for specific tasks. (huggingface.co, cloud.google.com) The company has been laying the plumbing for mobile use for months. Google’s Gemma mobile documentation says developers can deploy Gemma models on Android and iOS through the MediaPipe Large Language Model Inference application programming interface, which wraps on-device text generation. (ai.google.dev) Gemma 4 also arrives as Apple, Google, and Qualcomm push more artificial intelligence work onto chips inside phones and laptops, where latency is lower because data does not have to make a round trip to a remote data center. Google’s DeepMind page describes Gemma as a collection of lightweight open models built from the same technology behind Gemini. (deepmind.google) For people who want a chatbot, coding helper, or image-aware assistant without a monthly bill or constant internet connection, the smaller Gemma 4 models move more of that work to the device already in their pocket or on their desk. (blog.google, ai.google.dev)