Google DeepMind: Gemini, AlphaFold and AI research

Google DeepMind is Google’s AI research and development unit formed by bringing DeepMind together with the Brain team from Google Research. Its work spans general-purpose AI models and specialised scientific systems. This guide explains the organisation, the difference between its major technologies, and why a research breakthrough should not be read as a universal product guarantee.

By Nicholas Thackray. Sources checked 16 September 2026. This is a source-based reference profile, not a hands-on review.

What is Google DeepMind?

Google announced the combined unit on 20 April 2023, naming Demis Hassabis as its chief executive. The announcement distinguished Google DeepMind from Google Research, which would continue work across a wider range of computer-science subjects. Google DeepMind should therefore not be used as a catch-all name for every Google research project.

The lab’s company history dates the original DeepMind organisation to 2010 and Google Brain to 2011. Those dates describe the predecessor teams. The 2023 combination is a separate organisational event, not the founding date of the original DeepMind lab. The same current history page identifies Hassabis as chief executive.

For the broader industry context, explore Nuvastra’s AI company directory. The AI Fundamentals guide introduces the technical ideas behind the organisation’s work.

Gemini: models and the applications around them

The Gemini model page presents a family of systems with capabilities spanning different forms of information, including text, images, audio and video. It also links routes for trying Gemini and building with the models. The model, the consumer application and the developer environment are different things: each has its own interface, available features and conditions of access.

This distinction matters when an announcement describes a capability. A model demonstration does not establish that the feature is available in every app, region, account or API. A useful comparison records the exact version and access route before discussing performance or price. This company guide does not attempt to turn a changing product page into a permanent specification sheet.

AlphaFold: a specialised scientific system

AlphaFold concerns molecular structure prediction, rather than general conversational assistance. Google DeepMind’s overview separates AlphaFold Server, the AlphaFold Protein Structure Database and the AlphaFold 3 model code and weights. These resources serve different purposes: running a prediction, consulting existing predictions and obtaining model materials are not the same form of access.

The distinction is useful beyond biology. An AI research result may become a database, a hosted service or downloadable software, with different terms and practical requirements for each. Readers should identify which resource a claim concerns and follow the corresponding documentation. Success on a specialised scientific task should not be transferred to an unrelated model or interpreted as evidence that every output is correct.

From research result to usable system

Google DeepMind’s history includes work on game-playing systems, reinforcement learning and scientific applications. Its 2023 formation announcement set out an ambition to connect research with Google’s computational resources and products. That is a statement of organisational direction; it does not remove the need to examine individual results.

When reading a benchmark, ask which system was tested, what information it received, what counted as success and who carried out the evaluation. Keep vendor-reported results distinct from independently reproduced work. A score on a defined task is evidence about that task under those conditions, not a universal measure of intelligence.

The same discipline applies to a demonstration. An edited example can show that an outcome is possible without establishing how frequently it occurs, how much it costs or which failures were excluded. Nuvastra’s AI glossary explains the language of evaluations, agents and other concepts used in such claims.

How this profile should be used

Use this entry to identify the organisation and find the relevant primary documentation. Use an exact model record for changing specifications, and a dated research paper or evaluation report for performance claims. Comparing companies is useful for understanding their roles; comparing actual systems requires a more precise unit of analysis.

This page does not provide a live Gemini price list, a complete research catalogue or independently measured benchmark results. Source review: 16 September 2026. No independent product testing is claimed.