Google DeepMind: Gemini and Scientific AI Research

Google DeepMind combines model development with scientific AI research. Explore its 2023 formation, Gemini releases and leadership changes.

kindResearch lab
foundedApr 20, 2023
part ofGoogle

Latest

Updated

A research lab within Google

Google DeepMind was formed on April 20, 2023 by combining DeepMind and the Brain team from Google Research. Google's announcement described a new unit backed by Google's compute resources, while Google Research continued separately. The lab and its parent therefore have different roles in the Atlas.

Its portfolio includes Gemini models, open-weight Gemma models, media generation, agents and scientific systems such as AlphaFold. These research and model artifacts feed products and developer services within Google's larger business. A listed research objective should not be read as evidence that the objective has been achieved.

Leadership and model access changed over time

In August 2026 Google reorganized the lab's leadership. Demis Hassabis became its chair and Alphabet's chief scientist, while Koray Kavukcuoglu became senior vice president of Google DeepMind, reporting to Sundar Pichai. The announcement placed Gemini development, research, application and developer teams under Kavukcuoglu. These titles supersede the original formation announcement.

The September 2026 Gemini 4 Argon release began with staged access for trusted cyber defenders under Google's Fairwind program. That was a restricted deployment stage, rather than evidence that every Gemini user received the model or could download its weights.

Sources

  1. [1]
  2. [2]
    About Google DeepMind

    Google DeepMind · retrieved Oct 23, 2026

  3. [3]
  4. [4]

Google DeepMind history

2026

│
Google releases Gemini 4 Hydrogen

Google released Gemini 4 Hydrogen, a faster Gemini 4 model for agentic and everyday work that it calls a major upgrade from Gemini 3.8 Flash, at $1.50 and $7.50 per million input and output tokens. It reached the Gemini API, the Gemini app and Google Search for all users the same day.

Google
Model
│
Google releases Gemini Omni 1.2 Flash with 15-second video

Gemini Omni 1.2 Flash generates clips of 3 to 15 seconds, up from 10 seconds in Omni 1.1, with spoken dialogue and lip sync, at $0.07 per second of 720p video, 30 percent less than before. It is generally available on the Gemini API and Google Cloud.

Google
Model
│
Google DeepMind opens a Genie 4 research preview

Google DeepMind began a trusted-tester preview of Genie 4, a real-time world model with natural-language steering and native audio. Access for Google AI Ultra subscribers and Gemini API partners was still forthcoming.

Google
Model
│
Biohub's Virtual Biology Initiative grows to about $1.8 billion with US and AI lab money

With Biohub, United States Department of Energy, Meta, Isomorphic Labs

Biohub's Virtual Biology Initiative grew to about $1.8 billion in funding, data, computing and instruments, including more than $500 million over five years from the Energy Department and $300 million from Google DeepMind, Isomorphic Labs and Meta. Biohub's own April pledge and existing NIH data make up much of the rest.

Biohub
Deal
■
Google releases Gemini 4 Argon, first to vetted cyber defenders

Google's new frontier model, with a unified 1 million token input-output limit, went first to partners in its Fairwind cyber-defense program while Google took part in the US government's voluntary pre-release access process. On October 13 it reached the paid Gemini API, at an introductory $2 and $10 per million tokens until November 30, and AI Pro and Ultra subscribers in the Gemini app.

Google
Model
│
Google DeepMind releases AlphaGenome Atlas, mapping every human DNA variant

A roughly 1-petabyte dataset of precomputed molecular-effect predictions for all 9 billion possible single-letter changes in the human genome, built by running AlphaGenome across the whole genome rather than one variant at a time. Open for non-commercial research through a free web portal and API; DeepMind says it is more than 30 times the size of the AlphaFold Database.

Google DeepMind
Research
│
Google launches Gemini 3.8 Flash and a cyber-defense variant

Its third Flash-tier release in six weeks, plus 3.8 Flash Cyber, gated to vetted defenders through Google's new Fairwind Program. Google says the cyber variant wrote 2.6 times more correct Chrome vulnerability patches than rival models in its own tests.

Google
Model
│
Gemma passes a billion downloads

Google reported more than 100,000 community variants of Gemma two years after launch. Teams at NASA, Satlyt and Starcloud used the models for onboard image analysis, and the Wild Dolphin Project used them for dolphin-vocalization research.

Google
Open source
■
Hassabis moves to chair and Jeff Dean leaves Google

With Google

Koray Kavukcuoglu took over Gemini, the app, and frontier research as senior vice president; Demis Hassabis became chair of Google DeepMind and chief scientist of Alphabet; and Jeff Dean left after 27 years to start a public benefit corporation with Sanjay Ghemawat.

Google
Company
■
Lab employees ask Washington for a brake they can pull later

With OpenAI, Anthropic, Meta

More than a thousand people inside the frontier labs, including Anthropic's chief executive and OpenAI's chief scientist, asked the United States to back an international effort to build the tools for deliberately pacing automated AI development. The statement does not ask anyone to slow down now.

Pacing the Frontier
Culture
│
Gemma 4 drops the vision encoder

Google's open-weights model went encoder-free and multimodal in one architecture, at a size that runs on a single consumer card. Open weights from US labs had become a rarity worth noting.

Google
Model
│
Gemini Spark runs on Google's machines, not yours

A personal agent on a dedicated cloud VM, built on the Antigravity harness, working through Gmail, Docs and Slides with the laptop closed. Google's answer to the coding agents, aimed at everyone who does not write code.

Google
Product
■
Gemini Omni generates video and calls it a world model

Any input to any output, starting with video, with edits that build on one another while physics and characters hold across them. Google framed it as the move from predicting text to simulating reality, which is a large claim for something that launched inside YouTube Shorts.

Google
Model
■
Gemini 3.5 leads with the small model

Google shipped the Flash tier first and claimed it beat Gemini 3.1 Pro on coding and agentic work while running four times faster, at 76.2 percent on Terminal-Bench 2.1. Pro was promised for the following month, which is a new order of operations for a frontier release.

Google
Model
│
Google, Microsoft and xAI hand models to the government before release

With Center for AI Standards and Innovation, Microsoft, xAI

The Center for AI Standards and Innovation signed pre-deployment evaluation agreements with three more labs, extending the arrangement OpenAI and Anthropic made in 2024. Five weeks later the same government stopped a model rollout outright.

CNBC
Policy

2025

■
Gemini 3 ships into Search on day one

Google released Gemini 3 Pro in preview and put it into AI Mode in Search, the Gemini app, developer tools and Vertex AI on launch day. Its published scores covered different tool conditions and evaluation harnesses, so they describe specific tests rather than a universal ranking.

Google
Model
│
Google DeepMind ships Gemini Robotics 1.5

The vision-language-action model lets a robot reason through multi-step tasks and explain its steps before acting, with skills that transfer across robot bodies. Gemini Robotics-ER 1.5, the planning layer above it, opened to developers in the Gemini API the same day.

Google DeepMind
Model
■
Google hires Windsurf's CEO after OpenAI's acquisition collapses

With Windsurf, OpenAI

OpenAI's $3 billion deal for the AI coding startup fell apart over an IP dispute with Microsoft, so Google DeepMind paid $2.4 billion instead for a nonexclusive license to Windsurf's technology and to hire CEO Varun Mohan, a co-founder, and top researchers.

TechCrunch
Deal
│
AlphaEvolve improves on Strassen after 56 years

A Gemini-driven search over programs found a way to multiply 4x4 complex matrices in 48 scalar multiplications, one fewer than the 1969 record. The same system clawed back 0.7 percent of Google's worldwide compute by rewriting a data center scheduler.

Google DeepMind
Research

2024

│
Google DeepMind announces Veo 2, a sharper video model

Veo 2 generated video up to 4K resolution and several minutes long, with a better grasp of camera lenses, physics and cinematographic language than its predecessor. Google opened a waitlist for it inside the Labs tool VideoFX.

Google
Model
■
Google unveils Gemini 2.0 for what it calls the agentic era

Gemini 2.0 Flash beat the older 1.5 Pro on key benchmarks at twice the speed, adding native image and audio output and built in tool use. Google paired the launch with research previews of an agent that operates a browser and an upgraded universal assistant.

Google
Model
│
Google DeepMind's AlphaQubit learns to decode quantum computing errors

A transformer-based neural network trained to read the consistency checks in a quantum processor and predict where errors occurred, published in Nature alongside Google Quantum AI. It beat the best previous decoders on real data from Google's Sycamore chip and in simulations of larger devices.

Google
Research
│
Gemini 1.5 Flash and Project Astra debut at Google I/O

With Google

Google's answer to GPT-4o's launch day: Gemini 1.5 Flash for cheap high-volume work, a two million token context window, and Project Astra, a camera-and-voice assistant demo. Flash became the workhorse; Astra stayed a demo.

Google
Model
│
Google DeepMind and Isomorphic Labs introduce AlphaFold 3

With Isomorphic Labs

Google DeepMind and Isomorphic Labs introduced AlphaFold 3 for predicting biomolecular complex structures and launched a hosted server for noncommercial academic research. The release extended prediction beyond proteins; it did not publish model weights that day.

Isomorphic Labs
Model