Google consolidates AI under Gemini, centers on Gemini 3.5 and 1M‑token contexts

Google has unified multiple AI products and models under the Gemini label, with the Gemini 3.5 series as the current baseline, a multimodal Omni video capability, and family‑wide support for at least one‑million‑token context windows.

Google has consolidated its AI products and core models under the Gemini brand, making Gemini the umbrella for a multimodal model family plus several integrated product experiences. The rollout centers on the Gemini 3.5 series as the current baseline while older Gemini variants remain available; the rebrand also covers the chatbot formerly known as Bard, a planned replacement for Google Assistant on Android and Wear, Workspace integrations, and other branded tools that run on the same underlying models. This alignment matters because it unifies development, productization, and API access around a single model family.

At a technical level, Gemini models are multimodal large language models that natively accept and generate text, images, audio, video and code. Google says the family uses transformer architectures with standard pretraining and fine‑tuning workflows. For very large models, the company has shifted toward a mixture‑of‑experts design to improve compute efficiency at high parameter counts. Gemini Omni specifically extends Gemini’s multimodal scope to video: it can generate video from text, image, audio and video prompts, enabling creators and applications to combine different media as both inputs and outputs. That same multimodal primitive underpins features such as image description, chart interpretation, menu translation, and the generation of graphs or visualizations from supplied data.

A defining capability across the Gemini family is very long context support — at least a one‑million‑token window — which Google pioneered and continues to emphasize. That scale lets a single prompt ingest multiple long documents or large knowledge bases, changing workflows for tasks like contract analysis and multi‑document question answering by reducing the need for external retrieval or repeated context stitching.

Google exposes Gemini’s abilities both inside its own apps and devices and to third‑party developers via APIs, allowing external applications to leverage the same multimodal and long‑context features. The company reports state‑of‑the‑art results across the family and highlights its multimodal and long‑context strengths; however, competitors have been narrowing performance and capability gaps and public disclosure of internal architecture and training details remains limited.

For builders, the practical tradeoffs are concrete: the 1M‑token windows enable simpler retrieval‑augmented generation (RAG) patterns and direct ingestion of long documents, but fully using that context in production can substantially raise API costs. The mixture‑of‑experts architecture offers efficiency at scale, and Gemini’s multimodal primitives let developers combine text, image, audio and video in single requests — all factors teams must weigh when designing pipelines and estimating runtime expense.