Google Releases Multimodal Gemma 4 31B Model

Google DeepMind has expanded its open-weights lineup with the release of Gemma 4 31B IT, a powerful new multimodal model. This 31-billion-parameter release marks a significant milestone for the Gemma family, introducing vision-language capabilities at a substantial scale.

As a Vision Language Model (VLM), Gemma 4 31B can interpret and process information from both text and images simultaneously. This allows it to handle tasks that were previously out of reach for its text-only predecessors, such as describing the contents of a photograph or answering questions based on visual information.

The "IT" in the model's name signifies that it is instruction-tuned, meaning it has been specifically optimized to follow user prompts and perform conversational tasks. This fine-tuning makes it more reliable and useful for building interactive applications.

The introduction of a capable VLM under the Gemma name advances Google's open-source AI strategy, providing developers with a strong foundation for building sophisticated multimodal applications. The model is available under the Gemma license and can be accessed by researchers and developers on its Hugging Face repository.