Google Cloud unveils a massive package of artificial intelligence updates

In March 2026, Google Cloud unveiled a comprehensive update to its artificial intelligence ecosystem, encompassing new multimodal solutions, cost-effective models for content generation, and enhanced cybersecurity tools. This release aims to provide developers and enterprise customers with flexible tools for building scalable applications. An important part of the March announcements was not only the introduction of new generative neural networks but also the publication of detailed practical guides that help engineers optimize model training processes and configure infrastructure for specific business tasks.

A central element of the Vertex AI platform update was the launch of Gemini Embedding 2, Google's first native multimodal embedding model. This tool, currently available in public preview, addresses the complex task of unifying various data formats into a single coordinate system. Gemini Embedding 2 projects text information, images, video files, audio recordings, and complex documents into a unified embedding space. This approach enables cross-modal search and classification of media files without the need for separate models for each data type, although the exact dimensionality parameters of this unified vector space are not disclosed in the published materials.

To address tasks related to dynamic visual content creation, the Veo 3.1 Lite model was introduced, completing the lineup of generative solutions in the Veo 3.1 family. This neural network is specifically designed for creating video applications with high query volumes, offering developers flexibility in choosing tools based on their current needs. The main advantage of Veo 3.1 Lite is its cost-effectiveness, as generation costs are reduced by more than fifty percent compared to the Veo 3.1 Fast version, while query processing speed remains at the same level. To accelerate adoption, Google also released a specialized guide on prompt engineering for the entire Veo model series.

Significant improvements also affected the audio processing domain thanks to the update of the Gemini 3.1 Flash Live model. Google engineers enhanced the overall performance quality of the system, making it a more reliable foundation for enterprise customers and developers building AI voice agents. These agents are now capable of performing complex multi-step tasks at an industrial scale. As confirmation of the increased efficiency, a result of 90.8 percent is cited on the specialized ComplexFuncBench Audio benchmark, which assesses the model's ability for multi-step function calling considering various strict constraints, surpassing the performance of previous system generations. However, specific metrics of previous versions for direct comparison are absent from the report.

Amidst the development of generative technologies, Google Cloud announced the official completion of its acquisition of the Wiz platform, specializing in cloud and artificial intelligence security. The Wiz team is fully integrating into Google Cloud's structure, but the Wiz brand itself will be retained. The integration of these technologies aims to redefine approaches to data protection in the era of AI development. The combined ecosystem will provide customers with a comprehensive platform for securing cloud and hybrid environments, and will significantly accelerate the processes of preventing, detecting, and responding to emerging cyber threats in corporate infrastructure.

As part of an educational initiative for the engineering community, the Google team published a comprehensive guide for working with Nano Banana 2 and Nano Banana Pro models. The creation of this document was preceded by weeks of intensive testing, during which specialists examined the extreme capabilities of neural networks across all sorts of usage scenarios. The guide reveals non-obvious mechanisms for achieving the best results in content generation. Specifically, developers are advised to use a refined formula for composing queries, which includes a combination of reference images, instructions on their interrelationships, and a description of the new scenario, allowing for maximum control over the final output.

Hardware resource optimization issues are thoroughly discussed in a new developer guide for training models using Ironwood Tensor Processors. The material, prepared by Lilian Yu from Product Strategy and Operations and Product Manager Liat Berry, describes five specific strategies within the JAX and MaxText ecosystems. These strategies are designed to help engineers increase the efficiency of training processes and achieve peak performance on specialized Ironwood hardware, thereby reducing the time and computational costs for preparing large-scale models.

The March update cycle concludes with a methodological guide on creating production-ready AI agents using managed MCP servers. As a practical example, the document examines the demo agent Cityscape, created using the Google ADK application development kit. This agent clearly demonstrates the process of transforming a simple text query for generating a Kyoto cityscape into a unique image created by artificial intelligence. Although the guide details the agent's operational logic, it does not specify the system requirements for deploying development servers on the client side.