Gemma 4 Goes Open

Google released Gemma 4 on April 2 as a new family of open-weight artificial intelligence models under the Apache 2.0 license. (blog.google) The lineup includes four sizes: E2B, E4B, 26B A4B, and a 31B dense model. Google’s model card says the family handles text and images across the board, with audio support on the smaller models, and supports context windows up to 256,000 tokens in the larger variants. (ai.google.dev) Google said the 31B model ranked No. 3 among open models on the Arena AI text leaderboard on April 1, while the 26B model ranked No. 6. The company said those larger Gemma 4 models outperformed open models up to 20 times their size on that benchmark. (blog.google) An artificial intelligence model is a system trained to predict the next word, image patch, or token from patterns in large datasets. In practice, more parameters usually mean more capacity, but Gemma 4 is pitched around “intelligence per parameter,” or how much work a smaller model can do on less hardware. (blog.google) That hardware point is central to the release. Google said the smaller Gemma 4 models are designed for phones and laptops, while the 26B and 31B versions target consumer graphics cards, workstations, and cloud accelerators. (ai.google.dev) The licensing change is also new. Google’s open-source blog said Gemma 4 is the first Gemma release under the Open Source Initiative-approved Apache 2.0 license, which gives developers clearer terms for reuse, modification, and commercial deployment. (opensource.googleblog.com) Google is tying the release to a larger ecosystem push. The company said developers have downloaded Gemma models more than 400 million times and built more than 100,000 variants since the family launched in 2024. (opensource.googleblog.com) The release also adds features aimed at software agents, which are systems that call tools and complete multistep tasks instead of only answering prompts. Google’s model card lists native function calling, system prompt support, coding gains, and “agentic workflows” as core parts of Gemma 4. (ai.google.dev) Google is offering the models through its own cloud alongside downloads from Hugging Face and GitHub. Its cloud division said customers can deploy Gemma 4 on Vertex AI, fine-tune the 31B model on training clusters, and run inference on Cloud Run with NVIDIA RTX PRO 6000 graphics processors. (cloud.google.com) Gemma 4 follows a fast release cycle for Google’s open model line. Google’s release log shows Gemma launched in February 2024, Gemma 3 arrived in March 2025, and Gemma 4 was added on March 31, 2026, in E2B, E4B, 31B, and 26B A4B sizes. (ai.google.dev) The immediate effect is simple: developers now have a newer Google model family they can run, fine-tune, and ship with fewer licensing limits. The next test is whether Gemma 4’s 31B class can keep pace with closed systems once builders move from benchmarks to products. (blog.google)