Google DeepMind, the lab behind Gemini and Gemma

From DeepMind to Google DeepMind

DeepMind was founded in London in 2010 by Demis Hassabis, Shane Legg and Mustafa Suleyman, with the goal of building general-purpose AI systems. The lab made its name with AlphaGo, which beat Lee Sedol at Go in 2016, and later with AlphaFold in biology. Separately, Google Brain started in 2011 at X, Google’s moonshot division; its researchers published the 2017 paper “Attention Is All You Need”, which introduced the Transformer architecture behind today’s chatbots.

On April 20, 2023, Google combined the two teams into a single unit, Google DeepMind, led by Demis Hassabis. Its historic base remains in London’s King’s Cross district.

Two families, two policies

• Gemini: the main family, multimodal by design (text, images, audio and video). It is closed: the weights are not published. It includes a flagship reasoning model, faster Flash and Flash-Lite versions, and specialized models for voice, images, video and music.
• Gemma: open-weight models in several sizes, built to run anywhere from cloud servers to phones. You can download them, adapt them through fine-tuning and host them yourself.

This dual approach puts Google among the “mixed” labs: closed for its most powerful models, open for a lighter range.

How to access the models

• The Gemini assistant on the web and mobile: see the Gemini page.
• Google AI Studio and the Gemini API, for developers.
• Google Cloud, for businesses that need contractual terms and hosting options.
• Gemma: download from Hugging Face or Kaggle, or run it through local tools such as Ollama.

What sets Google DeepMind apart

In practice, its biggest advantage is distribution: Gemini models show up in products people use every day, from Search to Android and Workspace. The other distinctive trait is the scale of its scientific research, well beyond chatbots. To compare open models side by side, see the open source models page.