Google Cloud announced on the 8th that it has updated the models and enterprise features of the Vertex AI enterprise AI platform. Google Cloud officially released the Jeminae 1.5 Flash model, which was previewed back in May. Jeminae 1.5 Flash can be widely used in various large-scale cases, from retail company chat agents to document processing and research agents that can analyze entire data through combining short latency, competitive pricing, and a groundbreaking 1 million token context window.
Jeminae 1.5 Flash offers significant advantages over similar models like GPT-3.5 Turbo in terms of performance, latency, and cost-effectiveness. It provides a context window that is about 60 times longer than GPT-3.5 Turbo, offering up to 4 times cost-saving with context caching feature when input exceeds 32,000 characters.
Moreover, Jeminae 1.5 Pro, which supports up to 2 million token context windows, has also been released as a stable version, supporting various multimodal cases. Google Cloud supports enterprises in utilizing Jeminae 1.5 Pro for innovative purposes in diverse areas such as finding bugs in numerous code lines, searching for necessary information in research data, and analyzing hours of audio and video content.
To support customers in efficiently utilizing the extensive context windows of Jeminae 1.5 Pro and Jeminae Flash models, Google Cloud provides the context caching feature as a preview version. Vertex AI context caching helps customers save up to 76% on costs by utilizing cached data of frequently used contexts.
In order to increase trust in data and reduce hallucinations, enterprises need to ground model results based on strict enterprise standards such as data governance and data sovereignty, while also grounding model outcomes based on web, in-house, and third-party data and facts.
At Google I/O in May, following the official release of the Grounding with Google Search function of Vertex AI, Google Cloud announced a new service that allows enterprises to ground their generative AI agents with data from specific providers starting in the third quarter of this year. This enables businesses to integrate third-party data into generative AI agents, uncover unique use cases, and derive more enterprise knowledge throughout their AI experiences.
For data-intensive industries such as finance, healthcare, and insurance, where generative AI often needs to generate responses based solely on provided context, Google Cloud introduced the Grounding with High Fidelity mode in an experimental preview version. This mode supports grounding cases such as summarizing multiple documents, extracting data across financial data, and processing predefined document sets. Powered by fine-tuned Jeminae 1.5 Flash for generating answers using only provided context, it ensures high factual accuracy.
Google’s latest image generation model, Imagen 3, has been enhanced in various aspects compared to Imagen 2, providing superior image quality with over 40% faster image generation for prototyping and iterative tasks. It also offers improvements in prompt understanding, command execution, creation of lifelike group portraits, and control over text rendering within images.
Google Cloud offers a curated range of its own models, open models, and third-party models within Vertex AI. Recently, it added the newly released Cloud 3.5 Sonata by Anthropic to Vertex AI, allowing testing and deployment in the Google Cloud environment. In the second half of this year, Mistral Small, Mistral Large, and Mistral Codestral models will be provided in the Vertex AI model garden.
Additionally, based on the same research technology used in developing the Jeminae models, Google Cloud has introduced the Gemma follow-up models, a lightweight state-of-the-art open model product line. Gemma 2, officially released globally targeting researchers and developers, is available with parameter sizes ranging from 9 billion to 27 billion, offering significantly improved power, efficiency, and enhanced built-in security features. Gemma 2 can be used with Vertex AI starting this month.