NOOSIGMA AI / TRIAD / 03

Google and Gemini: who it is, history, products and architecture

Gemini is Google’s family of AI models, developed by Google DeepMind. This article covers where it comes from, how it evolved to Gemini 4 Argon, where it shows up in Google products and what the company says about the architecture.

  • Google DeepMind since April 2023
  • Multimodal from the start
  • Gemini 4 Argon on 13 Oct. 2026

Reviewed on

01

Who it is

Gemini is developed by Google DeepMind, which was formed in April 2023 by merging Google Brain and DeepMind. DeepMind was founded in the United Kingdom in 2010 by Demis Hassabis, Shane Legg and Mustafa Suleyman and acquired by Google in 2014. Demis Hassabis leads Google DeepMind.

The lab is known for scientific work such as AlphaGo, which beat champion Lee Sedol in 2016, and AlphaFold, which predicts protein structures. Hassabis and John Jumper received half of the 2024 Nobel Prize in Chemistry for that research. The transformer, the basis of today’s models, was also introduced by Google researchers in 2017.

Unlike the other two companies, Google takes Gemini into an ecosystem that already existed: Search, Android, Chrome, Workspace and Google Cloud, plus its own AI chips, the TPUs.

02

Timeline

  1. 2010–2014

    DeepMind is founded (2010) and acquired by Google (2014).

  2. 2016–2020

    AlphaGo beats Lee Sedol in 2016; AlphaFold stands out in protein prediction competitions in 2018 and 2020.

  3. 2017

    Google researchers introduce the transformer in the paper “Attention Is All You Need”.

  4. Apr. 2023

    Google Brain and DeepMind merge into Google DeepMind.

  5. May 2023

    Gemini is announced at Google I/O as the successor to PaLM 2.

  6. Dec. 2023

    Gemini 1.0, in three sizes: Ultra, Pro and Nano. Gemini Pro starts powering Bard.

  7. Feb. 2024

    Bard is renamed Gemini (8 February). Gemini 1.5 Pro arrives with mixture of experts and context of up to 1 million tokens. Google also releases Gemma, a family of open models.

  8. 2025

    Gemini 2.0 (January and February), Gemini 2.5 Pro (March), Gemini CLI (June) and Gemini 3 Pro (November, in preview), followed by Gemini 3 Flash (December).

  9. 2026

    Gemini 3.1 Pro (February), Gemini 3.5 Flash (May) and new Flash versions between July and September.

  10. Oct. 13, 2026

    Gemini 4 Argon, the first model of the 4 family, is announced and becomes available in the API, the Gemini app and Search for eligible subscribers.

03

Products

Gemini app

Google’s assistant on web, Android and iOS, with Google AI Pro and Ultra plans.

Search (AI Mode) and Chrome

Generated answers and AI features built into search and the browser.

Workspace and Gemini for business

AI in email, documents and meetings. In October 2026, Google Cloud presented an agent that brings research, creation and code together with access, audit and cost controls.

Android and devices

Gemini Nano runs on the device, from phones such as the Pixel 8 Pro and from partners such as Samsung.

Gemini API, AI Studio and Google Cloud

Access for developers and companies. Google Cloud’s Model Garden is also set to receive OpenAI’s models, per an October 2026 announcement.

Gemma and Gemini CLI

Smaller open models and an open-source agent for the terminal.

Google Antigravity

A development environment cited in the Argon launch, where subscribers and companies can use the model.

Nano Banana

A line of Gemini-based image generation and editing models.

04

Architecture and training

Gemini models are decoder-only transformers optimized for Google’s TPUs. The 1.0 generation used multi-query attention, a technique that lowers the cost of running the model.

The project’s hallmark is native multimodality: from the start, the model was designed to process text, images, audio, video and code together, instead of joining separate models. For audio, sound is sampled at 16 kHz and converted into tokens by the Universal Speech Model.

From Gemini 1.5 Pro (2024), Google began using a mixture-of-experts architecture, with context windows of up to 1 million tokens in production. Gemini 3 Pro (2025) was described as a sparse mixture-of-experts, natively multimodal reasoning model.

About Gemini 4 Argon, Google says training aimed at respecting system boundaries, reporting negative findings honestly and asking for human confirmation before irreversible steps, using adversarial training and a misalignment monitor during reinforcement learning. The API offers a unified window of 1 million input and output tokens.

05

What to look at

The declared advantage is integration: anyone who already works with Google tools finds Gemini in Workspace, Search and Android. Argon’s availability depends on plan and product, and the introductory price given changes from 30 November 2026.

As with any model, performance numbers released at launches describe specific conditions. Test with your own files and tasks, and check the official page of the product you plan to use.

06

Sources