Google Releases Gemma 4: Four Variants, Apache 2.0 License, Fully Covers Everything From Mobile to Servers

Google DeepMind officially released Gemma 4 on April 6. This is the fourth generation of the Gemma series open-weight models. It comes in four versions, the license has been changed to Apache 2.0, and it now adds native support for multimodal inputs and agent workflows.

Four specifications: cover everything from mobile devices to servers

The 31B dense model ranks third among open-source models on the Arena AI text leaderboard; the 26B mixture-of-experts model ranks sixth. Google says that in Arena AI benchmark tests, the two large models still have an advantage over other models whose parameter counts are up to 20 times higher.

Core capabilities: multimodal, audio, and 140 kinds of languages

All four versions natively support multimodal inputs for video and images, covering OCR and chart understanding. The edge versions (E2B, E4B) additionally support audio input and can be used for speech recognition scenarios.

All models support function calling, structured JSON output, and native system instructions for building agent workflows. The training data covers more than 140 languages.

The license has changed from custom to Apache 2.0, making commercial use more friendly

The custom license used by past Gemma series imposed restrictions on modification and re-release. Gemma 4 adopts the industry-standard Apache 2.0 license, allowing commercial use, modification, and re-release without needing to sign a separate agreement with Google. Google says this makes Gemma 4 more suitable for enterprise and developer applications.

All four models are now fully available for use. Google has not yet announced a timeline for integrating the edge models into Android or consumer hardware products.

This article Google releases Gemma 4: four specifications, Apache 2.0 licensing, full coverage from phones to servers was first published in ABMedia Chain News.