Google AI: Gemma 4 & Veo 3.1

Google just pushed three separate changes into one weekly update, and together they redraw a basic developer choice: run an open model on your own hardware, buy a cheaper video model through an application programming interface, or wire Google’s hosted models straight into a mobile app with Firebase. The names in this round were Gemma 4, Veo 3.1 Lite, and Firebase AI Logic. (blog.google 1) (blog.google 2) (firebase.google.com) An open model is like getting the engine instead of renting the whole car: you download the model weights, run them on your own machines, and decide where the data lives. Google says Gemma 4 is its new open family, released under the Apache 2.0 license, with sizes from Effective 2B and Effective 4B up to 26B Mixture of Experts and 31B Dense. (developers.googleblog.com) (blog.google) Gemma 4 is not just a text bot. Google’s model card says the family is multimodal, takes text and image input, supports audio on smaller models, works in more than 140 languages, and stretches to a 256,000-token context window, which is roughly the difference between reading a paragraph and keeping a short book in memory. (ai.google.dev) Google is aiming Gemma 4 at phones, laptops, and edge devices where sending every request to a cloud server is too slow, too expensive, or too sensitive. Google AI Edge says Gemma 4 can be used for multi-step planning on-device, and Google’s main launch post says the family is built to handle logic and agent-style workflows instead of only simple chat. (developers.googleblog.com) (blog.google) The second piece is video generation, which is one of the most expensive corners of generative artificial intelligence because every answer is a moving sequence of images with timing, motion, and often sound. Google introduced Veo 3.1 Lite as a cheaper tier and said it costs less than half of Veo 3.1 Fast while running at the same speed. (blog.google) (ai.google.dev) That price cut changes what teams can afford to build. Google describes Veo 3.1 Lite as a model for high-volume apps, and its Gemini application programming interface docs position it for text-to-video generation, image-to-video generation, editing, and cinematic controls, which is the difference between making one polished demo clip and generating thousands of clips inside a product. (blog.google) (ai.google.dev 1) (ai.google.dev 2) Google also folded these launches into the same Gemini application programming interface timeline. The Gemini changelog lists Veo 3.1 Lite Preview on March 31, 2026, new Flex and Priority inference tiers on April 1, 2026, and Gemma 4 model releases on April 2, 2026, which shows Google bundling model launches and pricing controls into one developer pipeline instead of treating them as separate products. (ai.google.dev) The third piece is less flashy and more practical: Firebase AI Logic is now the migration path for apps that used to rely on Vertex AI in Firebase. Google’s migration guide says the old software development kits could only use the Vertex AI Gemini application programming interface, while the new Firebase AI Logic software development kits let developers choose either the Gemini Developer API or the Vertex AI Gemini API. (firebase.google.com) That choice matters because mobile teams often want to ship one feature across Apple, Android, web, Flutter, and Unity without building a full server stack first. Firebase says its client software development kits are available for Swift, Kotlin, Java, JavaScript, Dart, and Unity, and are designed for direct use in mobile and web apps with Firebase security features layered around them. (firebase.google.com 1) (firebase.google.com 2) Put together, the update gives developers three different levers for the same problem. If they want control and local execution, Google is pushing Gemma 4; if they want cheaper generated video at scale, Google is pushing Veo 3.1 Lite; if they want the fastest path from prototype to shipped app, Google is pushing Firebase AI Logic and a cleaner migration path into the Gemini ecosystem. (blog.google) (blog.google) (firebase.google.com)