Google makes Gemini generally available on Vertex AI
Google has launched its Gemini family of models (Ultra, Pro, Nano) on its Vertex AI platform, with Gemini Pro now generally available for enterprise workloads. The move positions Gemini as a direct competitor to OpenAI and Anthropic for enterprise customers. Gemini is also being highlighted for its multimodal capabilities, including built-in OCR for processing scanned documents.
- Vertex AI provides a suite of MLOps tools designed to manage the entire lifecycle of Gemini models, including Vertex AI Pipelines for workflow orchestration, a Model Registry for versioning, and a Feature Store that now supports vector embeddings. - For model customization, Vertex AI supports various tuning techniques beyond prompt design, including adapter-based methods like LoRA (Low-Rank Adaptation), which allows for more efficient fine-tuning of models for specific tasks. - Alongside the generally available Gemini 1.0 Pro, Google has introduced Gemini 1.5 Pro in preview on Vertex AI, which features a breakthrough 1 million token context window, enabling the processing of large documents, hour-long videos, or codebases with over 30,000 lines in a single prompt. - The pricing for Gemini Pro on Vertex AI follows a pay-as-you-go model based on token consumption, with separate rates for input and output tokens, allowing for more granular cost management of inference pipelines. - The platform is built for enterprise compliance and data governance, allowing customization of Gemini with full data control and meeting standards like SOC2 and HIPAA. - Early enterprise adoption cases highlight specific uses; for instance, KPMG deployed Gemini Enterprise to 90% of its workforce within two weeks for innovation and productivity, and Gordon Food Service is using it to support its IT and developer teams. - In the competitive landscape, Gemini's key differentiator is its native multimodality, processing text, video, and audio simultaneously, whereas some enterprise users report that models from OpenAI and Anthropic currently have an edge for certain software development and data analysis use cases. - The underlying infrastructure for training and serving Gemini models on Vertex AI includes Google's latest custom Tensor Processing Units, like Cloud TPU v5p, which are designed for training large-scale generative AI models more efficiently.
