Google Presents Gemma 4: A New Generation of Open-Source AI Models Under Apache License
Google DeepMind proudly announced the launch of Gemma 4 on April 2, 2026. This marks their most ambitious open-source AI project to date. The new model family, available in four variants, ranges from 2 billion to a staggering 31 billion parameters. These models are designed to operate on a variety of devices, from smartphones to enterprise servers.
Built upon the robust research foundation of its predecessor, Gemini 3, Gemma 4 delivers advanced reasoning capabilities. It supports over 140 languages and offers a multimodal understanding of text, images, and audio. The models feature context windows of up to 256,000 tokens. They are optimized for agentic workflows and offline code generation, making them a versatile tool in any developer’s arsenal.
In a significant departure from previous releases, Google is distributing Gemma 4 under the commercially permissive Apache 2.0 license. This gives developers complete freedom to modify and deploy the AI for commercial purposes. The 31-billion parameter model ranks third on Arena AI’s text leaderboard, outperforming models twenty times its size.
The release of Gemma 4 comes as Google faces increasing competition from Chinese open-weight models and Meta’s Llama 4. Android developers can prototype applications through the AICore Developer Preview. This comes with day-one support for popular frameworks such as Hugging Face, vLLM, and Ollama.
Source: https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/