Google Releases Gemma 4 Multimodal Open Model

Google DeepMind has expanded its open-source offerings with the release of Gemma 4, a new generation of its popular model family. The initial variant, Gemma 4 26B-A4B Instruct, is a powerful multimodal model designed for conversational AI and assistant tasks.

Efficient by Design

Gemma 4 employs a sparse Mixture-of-Experts (MoE) architecture, a design choice that balances model scale with computational cost. While the model contains a total of 26 billion parameters, only 4 billion are active during inference for any given input. This approach allows Gemma 4 to achieve the performance associated with larger models while requiring significantly less processing power.

Key specifications include:

• Total Parameters: 26 billion
• Active Parameters: 4 billion
• Modalities: Text and Vision (VLM)
• License: Apache 2.0

This release is notable for its multimodal capabilities. As a vision-language model (VLM), Gemma 4 can understand and process both text and image inputs, making it suitable for a wide range of applications from image captioning to visual Q&A. This specific version is instruction-tuned, meaning it has been optimized to follow user prompts and engage in helpful dialogue.

By releasing Gemma 4 under the permissive Apache 2.0 license, Google continues to support the open-source AI community. This allows developers and researchers to freely build upon, modify, and deploy the model for both academic and commercial purposes, further accelerating innovation in the field.