Google introduced Gemma 4, a set of four open-weight models under the Apache 2.0 license. The license update from previous Gemma versions expands the available options for modification, reuse, and commercial use.

These models are designed to support various hardware, from mobile devices to server-scale inference. The announcement was made by Google DeepMind, focusing on reasoning, agent-based workflows, and multimodal input support.

Gemma 4 Model Sizes and Hardware Targets

Google announced that Gemma 4 has been released in several configurations. These include E2B, optimized for edge devices and targeting mobile hardware and low-power states; and E4B, optimized for similar devices but with higher capacity.

  • 26 billion parameter mixture of experts model, designed for server-scale use, currently ranks sixth in the Arena AI text leaderboard among open models.
  • 31 billion parameter dense model ranks third in the same leaderboard.

Google states that the 31B and 26B models performed 20 times better than other models with fewer parameters in the Arena AI benchmark. These performance claims are based on Google's own evaluations at the time of the models' release.

Capabilities of the Gemma 4 Model Family

All four Gemma 4 models come with native multimodal support for video and images at various resolutions; OCR and graphic understanding are among the main use cases. The E2B and E4B models also include native voice input for speech recognition.

Context window lengths go up to 128K tokens for edge models and 256K tokens for the 26B and 31B models. All models support function calls, structured JSON output, and native system instructions for creating agent workflows. Google notes that the models are trained natively in over 140 languages.

Transition from Proprietary Gemma License to Apache 2.0

Previous Gemma model versions used a proprietary license that restricted modification and redistribution. Gemma 4 now uses the Apache 2.0 license, offering standardized terms for commercial use, modification, and distribution; this does not require a separate agreement with Google. Google states that this change makes Gemma 4 more suitable for business and developer applications.

All four Gemma 4 models are now available. However, Google has not provided a timeline for when the edge models will be integrated into Android or consumer hardware products.