Google has announced the next-generation Gemini 1.5 as it expands access to Gemini 1.0 via Vertex AI

The Gemini models were first released to a select number of customers, including Samsung and Jasper, for them to build a better AI experience for their customers using Vertex AI. Now, Google has expanded the availability of Gemini 1.0 Pro and Ultra, making Pro generally available to all Vertex AI users and Ultra available to Vertex AI users via allowlist. Gemini 1.0 Pro is a scalable model offering an appealing balance between performance and cost for the most common AI tasks, such as content generation, summarization, and classification. In contrast, Gemini 1.0 Ultra excels at performing complex tasks, including coding, reasoning, and instruction.

In addition to the expanded availability of the Gemini 1.0 models, Google has also announced its next-generation Gemini 1.5. The mid-size Gemini 1.5 Pro is the first 1.5 model available in private preview on Vertex AI. Despite being a mid-size model, the Gemini 1.5 Pro boasts a performance comparable to the Gemini 1.0 Ultra. Moreover, Gemini 1.5 Pro features a breakthrough experimental capability: its 1-million-token context window enables processing massive amounts of information. The context window length approximately translates to 1 hour of video, 11 hours of audio, codebases with over 30,000 lines of code, or over 700,000 words. The model will initially ship with a 128K-token context window, but a selection of developers and enterprise customers will have access to the extended-length context window in the private preview.

The Gemini API in Vertex AI lets developers build the next generation of multimodal AI applications in a cloud-based platform incorporating the necessary tools to deploy and maintain Gemini-powered applications. Developers can use Vertex AI to customize the Gemini models Low-Rank Adaptation (LoRA), with upcoming techniques such as reinforcement learning from human feedback (RLHF) and distillation. Developers can profit from the generally available grounding and function calling to augment the Gemini models. Moreover, Vertex AI also provides tools for scaling and managing the models in production and minimal coding tools to create conversational and search assistants supported by the Gemini models.

The next-generation Gemini 1.5 models boast an increase in performance and efficiency supported by its new Mixture-of-Experts (MoE) architecture. Gemini 1.5 identifies the type of input it has been given and then activates the relevant pathways in its neural networks with the benefit of significantly increased efficiency. Gemini's improved architecture also enables faster learning of complex tasks, thus making the models more efficient to train and serve, which also translates into the capability of improving the models and delivering optimized versions faster than ever. In response to Gemini 1.5 being more capable than its predecessors, Google has committed to performing novel research on safety risks and conducting extensive safety evaluations, including tests accounting for Gemini 1.5's expanded context window length.

More details on Gemini's performance, evaluations, and safety testing are available in the original announcement and the Gemini 1.5 Pro technical report.