Google DeepMind
Last reviewed
Sources
66 citations
Review status
Source-backed
Revision
v10 · 8,561 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
66 citations
Review status
Source-backed
Revision
v10 · 8,561 words
Add missing citations, update stale details, or suggest a clearer explanation.
Google DeepMind is Alphabet's central artificial intelligence research and product organization, formed on April 20, 2023 by merging the DeepMind lab (founded in London in 2010) with Google Brain. It is the maker of the Gemini family of models and of AlphaFold, the protein-structure system whose work earned co-founder and CEO Demis Hassabis and director John Jumper the 2024 Nobel Prize in Chemistry. Headquartered at 6 Pancras Square in the King's Cross district of London, with major research outposts in Mountain View, New York, Zurich, Paris, Tokyo, Toronto, Tel Aviv, and Seattle, Google DeepMind develops general-purpose AI systems and conducts research spanning reinforcement learning, protein structure prediction, weather forecasting, mathematics, robotics, materials discovery, and large language models. The organization states that its mission is "to build AI responsibly to benefit humanity."[1][2][63]
DeepMind was founded in London in 2010 by Hassabis, Shane Legg, and Mustafa Suleyman. Google acquired the company on January 26, 2014 in what was then the largest European AI acquisition, for a reported figure between roughly $400 million and $650 million. After nine years operating semi-independently as an Alphabet subsidiary, the lab was merged with Google's internal Brain team to form a single, unified AI research and product organization. As of 2026, Google DeepMind sits at the center of Alphabet's strategy to compete with OpenAI, Anthropic, and other frontier model developers. Its Gemini family of multimodal models powers Google Search, Workspace, Android, and dozens of other Alphabet products, while its scientific projects, AlphaFold, AlphaProof, GNoME, and GraphCast among them, have produced some of the most significant applied AI results of the past decade.[3][4]
| Attribute | Detail |
|---|---|
| Type | AI research and product division of Alphabet |
| Formed | April 20, 2023 (merger of DeepMind and Google Brain) |
| DeepMind founded | 2010, London, by Demis Hassabis, Shane Legg, and Mustafa Suleyman |
| Acquired by Google | January 26, 2014, for a reported ~$400-650 million |
| CEO | Demis Hassabis, 2024 Nobel laureate in Chemistry |
| Headquarters | 6 Pancras Square, King's Cross, London |
| Flagship model family | Gemini (Gemini 3 Pro launched November 18, 2025) |
| Best-known science | AlphaFold, AlphaGo, GraphCast, GNoME, AlphaProof |
| Mission (stated) | "to build AI responsibly to benefit humanity" |
DeepMind Technologies Limited was incorporated in London in late September 2010 and announced publicly in November 2010. The three founders were Demis Hassabis, Shane Legg, and Mustafa Suleyman. Hassabis, a former child chess prodigy who had reached master standard at age 13 and co-designed the simulation game Theme Park at Bullfrog Productions while still a teenager, had returned to academia in the early 2000s to pursue a PhD in cognitive neuroscience at University College London under Eleanor Maguire, completing it in 2009. He held a Henry Wellcome postdoctoral research fellowship at the Gatsby Computational Neuroscience Unit at UCL, where he met Legg.[5][6]
Shane Legg, a New Zealander, had completed a PhD at the Dalle Molle Institute for Artificial Intelligence Research (IDSIA) in Lugano, Switzerland under Marcus Hutter on the foundations of machine super intelligence. Legg served as chief scientist of DeepMind and remains chief AGI scientist at Google DeepMind today. Mustafa Suleyman, a Syrian-British entrepreneur and policy specialist, had previously co-founded a Muslim youth telephone counseling service. Suleyman led the applied and policy operations of the new company.[5][7]
DeepMind's founding thesis combined two ideas: that artificial general intelligence (AGI) was the right north-star goal, and that it could be reached most efficiently by drawing inspiration from the biological brain. The slogan "solve intelligence, then use it to solve everything else" became a recurring formulation of the lab's mission, which Hassabis has described as the original two-step business plan: "Step one: solve intelligence. Step two: use it to solve everything else."[64] Early seed investment came from Founders Fund's Peter Thiel, Elon Musk, Skype co-founder Jaan Tallinn, Horizons Ventures, and Scott Banister. The company began with around fifteen employees occupying offices in Russell Square, central London, before moving north to King's Cross in 2018.[3][8]
In its initial years, DeepMind focused on deep reinforcement learning. The 2013 NIPS workshop paper "Playing Atari with Deep Reinforcement Learning" by Volodymyr Mnih and colleagues demonstrated that a single deep Q-network (DQN) architecture could learn to play multiple Atari 2600 games at human level directly from raw pixel input and the score, with no game-specific feature engineering. The follow-up Nature paper, "Human-level control through deep reinforcement learning," appeared in February 2015 and is widely credited with founding the modern field of deep RL. The DQN result was decisive in attracting Google's interest.[9]
On January 26, 2014, Google confirmed it had acquired DeepMind. Reported figures varied: many sources cited approximately $400 million USD, while others put the deal at around £400 million (roughly $500-650 million). The acquisition closed shortly after a competitive process in which Facebook had also pursued the company. According to multiple accounts, Google co-founder Larry Page personally drove the deal after a meeting with Hassabis, who had visited Mountain View earlier in 2013. The transaction was, at the time, the largest European AI acquisition on record.[10][11]
As part of the agreement, DeepMind negotiated several unusual concessions. The team was to remain in London. The company would not work on military applications. An ethics and safety review board, jointly staffed by representatives of Google and DeepMind, was to be created to oversee research with potential dual-use risks. The lab was also permitted to continue publishing in peer-reviewed journals and to maintain a strong academic culture, including conference attendance and open-source releases. Google declined to disclose the membership of the original ethics board, which became a subject of press speculation in the years that followed. In October 2017, DeepMind separately launched a public research unit called DeepMind Ethics & Society, distinct from the closed acquisition-era ethics board.[3][12]
The acquisition gave DeepMind access to Google's computational infrastructure, including its growing fleet of tensor processing units (TPUs), beginning with first-generation TPUs designed for inference workloads in 2015. In return, Google gained one of the world's most concentrated AI research teams. In 2016, DeepMind engineers Richard Evans and Jim Gao reported that the application of deep learning controllers to cooling units at one of Google's hyperscale data centers reduced cooling energy consumption by 40 percent, equivalent to a 15 percent reduction in overall power-usage effectiveness. The work was an early demonstration that DeepMind techniques could deliver concrete operational savings within Alphabet.[13]
AlphaGo, DeepMind's Go-playing program, became the public-facing breakthrough that changed the broader perception of AI. In October 2015, AlphaGo defeated the European Go champion Fan Hui 5-0 in a closed match in London. This was the first time a computer program had beaten a professional human Go player on a full 19x19 board without handicap. The result was published in Nature on January 27, 2016 in the paper "Mastering the game of Go with deep neural networks and tree search" by David Silver, Aja Huang, and colleagues, and was widely regarded as having occurred at least a decade ahead of expert predictions.[14]
The headline match took place between March 9 and 15, 2016, when AlphaGo faced 18-time world champion Lee Sedol in a five-game series at the Four Seasons Hotel in Seoul, South Korea, for a $1 million purse. AlphaGo won 4-1, with Lee securing a single victory in game four through an unexpected play known as "move 78" or the "hand of god" that briefly destabilized the AI's evaluation. Game two contained AlphaGo's celebrated "move 37," a shoulder-hit on the fifth line that initially appeared to be a mistake but was later judged a deeply original strategic stroke. The match was streamed live on YouTube and television worldwide, with an estimated 200 million viewers. The Korea Baduk Association awarded AlphaGo an honorary 9-dan professional ranking, and the prize was donated to charity.[14][15]
In early 2017, an upgraded version called "Master" played 60 fast online games against top professionals, winning every one. In May 2017, AlphaGo Master defeated the world's top-ranked player, Ke Jie, 3-0 in a three-game match at the Future of Go Summit in Wuzhen, China, after which DeepMind retired AlphaGo from competitive play. The October 2017 paper "Mastering the game of Go without human knowledge" introduced AlphaGo Zero, a version that learned entirely from self-play, with no human game data, and within a few days surpassed all previous incarnations.[14]
AlphaZero, described in a December 2018 Science paper, generalized AlphaGo Zero's tabula-rasa self-play approach beyond Go to chess and shogi (Japanese chess). Starting from random play and given only the rules, AlphaZero achieved superhuman performance in all three games within a single 24-hour training run on TPU hardware. It defeated Stockfish 8 (then the strongest specialized chess program) in a 100-game match without losing, defeated Elmo (a leading shogi engine), and decisively beat AlphaGo Zero in Go.[16]
In November 2019, DeepMind extended this line further with MuZero, a system described in the arXiv paper "Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model" and published in Nature in 2020. MuZero matched AlphaZero in Go, chess, and shogi but, crucially, did so without being told the rules, instead learning a latent dynamics model of the environment from interaction. The same algorithm achieved state-of-the-art results on the Atari Learning Environment, unifying model-based and model-free reinforcement learning in a single system.[17]
AlphaStar, introduced in January 2019 and detailed in an October 2019 Nature paper, extended DeepMind's game-playing research to the real-time strategy game StarCraft II. Unlike chess and Go, StarCraft II requires decision-making under partial information, with continuous time, large action spaces, and macro-economic and tactical reasoning intertwined. By using a population-based multi-agent reinforcement learning scheme called the AlphaStar League, the system learned to defeat top human players. By October 2019 it had reached Grandmaster level on the official competitive ladder across all three playable races (Terran, Protoss, and Zerg) under standard tournament conditions, becoming the first AI to reach the top league of a major esport without race or game restrictions.[18]
DeepMind's pivot toward applying AI to scientific problems began in earnest with AlphaFold, entered into the 13th Critical Assessment of protein Structure Prediction (CASP13) competition in late 2018, where the system placed first by a wide margin. AlphaFold 2, competing in CASP14 in November and December 2020, achieved a median Global Distance Test (GDT_TS) score of 92.4 across all targets, accuracy comparable to experimental methods such as X-ray crystallography. The CASP organizers and many in the structural biology community described the result as having effectively solved the long-standing single-domain protein structure prediction problem, a grand challenge dating back to Christian Anfinsen's Nobel-winning thermodynamic hypothesis from 1972.[19]
The AlphaFold 2 architecture, detailed in the July 15, 2021 Nature paper "Highly accurate protein structure prediction with AlphaFold" by John Jumper and colleagues, introduced an attention-based neural network module called the Evoformer that processes a multiple sequence alignment together with pairwise residue features, alongside a structure module that produces 3D coordinates iteratively. The complete AlphaFold 2 source code was released on GitHub under an Apache 2.0 license a few days later, accompanied by a structures paper by Kathryn Tunyasuvunakool and colleagues describing predicted structures for the entire human proteome.[19]
In July 2021, DeepMind launched the AlphaFold Protein Structure Database in partnership with the European Molecular Biology Laboratory's European Bioinformatics Institute (EMBL-EBI), initially containing predictions for around 365,000 proteins (including roughly 98.5 percent of the human proteome) and 20 model organisms. By July 2022 the database had grown to over 200 million predicted structures, covering nearly every catalogued protein sequence known to science. By the time of the 2024 Nobel announcement, more than 2 million users from over 190 countries had accessed AlphaFold predictions, and the system had become a standard tool in drug discovery, malaria research, antibiotic-resistance studies, plastic-degrading enzyme engineering, and basic biology.[20][65]
In August 2019, co-founder Mustafa Suleyman was placed on administrative leave from DeepMind following internal complaints about his management practices. The matter was investigated by an external law firm. According to subsequent reporting in The Wall Street Journal in January 2021, Suleyman was sidelined because he had bullied staff. He apologized publicly, saying he "accepted feedback that, as a co-founder at DeepMind, I drove people too hard and at times my management style was not constructive." In December 2019, Suleyman left DeepMind to take an AI policy role at Google. He departed Google entirely in early 2022 and co-founded Inflection AI with Reid Hoffman and Karen Simonyan in March 2022. In March 2024, after Microsoft acquired most of Inflection's staff and licensed its technology, Suleyman was named CEO of Microsoft AI.[21][22]
On April 20, 2023, Alphabet CEO Sundar Pichai announced that DeepMind and Google Brain would be merged into a single unit called Google DeepMind. In the announcement, Demis Hassabis wrote that "combining our talents and efforts will accelerate our progress towards a world in which AI helps solve the biggest challenges facing humanity," and framed the new unit as bringing "together our world-class talent in AI with the computing power, infrastructure and resources to create the next generation of AI breakthroughs."[23][63] The decision was driven by the launch of ChatGPT in November 2022, which had triggered what Pichai reportedly described internally as a "code red" for Google's competitive position in AI. Through 2022 and into 2023, Pichai held a series of leadership meetings to redesign Google's AI strategy, ultimately concluding that the duplication of effort between the two flagship AI groups was no longer sustainable. The two organizations had occasionally clashed over compute allocation, hiring, and credit, and DeepMind in particular had at times sought greater independence from Alphabet.[23][24]
Demis Hassabis was named CEO of the combined entity. Jeff Dean, the long-time head of Google AI and a co-founder of Google Brain, became chief scientist of both Google Research and Google DeepMind, reporting directly to Pichai. Koray Kavukcuoglu, previously vice president of research at DeepMind, was elevated to chief technology officer; in mid-2024 he was given the additional, newly created title of Google's chief AI architect. Eli Collins, formerly vice president of product at Google Research, joined Google DeepMind as vice president of product.[23][24][25]
The merger combined DeepMind's strengths in fundamental research, reinforcement learning, and scientific applications with Google Brain's expertise in scaled distributed training and productization. Google Brain had been the original home of the Transformer architecture, introduced in the 2017 paper "Attention Is All You Need" by Ashish Vaswani and colleagues, and of much of the early work on PaLM, T5, and the Pathways infrastructure. The unification gave the new organization end-to-end ownership of model research, training infrastructure, and product integration.[24]
The first major output of the merged organization was Gemini 1.0, announced on December 6, 2023 by Pichai and Hassabis at a virtual press conference. Gemini was Google's first natively multimodal flagship model, trained from the start on text, code, images, audio, and video tokens, rather than bolting modalities onto a text-only base. Three sizes were announced at launch: Gemini Ultra for highly complex tasks, Gemini Pro for general use, and Gemini Nano for on-device deployment on Pixel phones. Gemini 1.0 Pro was integrated into Bard on launch day, while Gemini Ultra became broadly available in February 2024 as the engine of "Gemini Advanced," a new tier of the Google One subscription. Bard itself was rebranded "Gemini" at the same time.[26]
Subsequent years saw a rapid cadence of Gemini releases, including the long-context Gemini 1.5 family in early 2024, the agentic Gemini 2.0 Flash in December 2024, the reasoning-focused Gemini 2.5 in March 2025, and Gemini 3 Pro on November 18, 2025. The Gemini 3 launch was accompanied by an unusual day-one rollout across Google Search AI Mode, the Gemini App, and Vertex AI, reflecting the increasing strategic centrality of the family.[27]
The day-to-day leadership of Google DeepMind is concentrated in a small group of executives, almost all of whom report into CEO Demis Hassabis. Hassabis in turn reports directly to Alphabet CEO Sundar Pichai and, according to a January 2026 CNBC interview, communicates with Pichai daily.[28]
| Role | Name | Background |
|---|---|---|
| CEO and Co-founder | Demis Hassabis | UCL neuroscientist, former game designer at Bullfrog and Lionhead, founder of Elixir Studios; 2024 Nobel laureate in Chemistry; knighted in 2024 |
| Chief Operating Officer | Lila Ibrahim | Former Coursera president and Intel executive; joined DeepMind in 2018, now Chief AI Readiness Officer at Google DeepMind |
| Chief Technology Officer / Chief AI Architect | Koray Kavukcuoglu | NYU PhD under Yann LeCun; joined DeepMind in 2012; named CTO January 2024, Chief AI Architect mid-2024 |
| Chief Scientist (jointly with Google Research) | Jeff Dean | Co-founder of Google Brain, designer of MapReduce, Bigtable, Spanner, and the original TensorFlow project |
| VP of Research / AI for Science | Pushmeet Kohli | Indian-born researcher, former MSR Cambridge; leads AI for Science and Strategic Initiatives Unit |
| Co-founder, Chief AGI Scientist | Shane Legg | IDSIA PhD under Marcus Hutter; long-time AGI theorist within DeepMind |
| VP of Product | Eli Collins | Joined from Google Research in 2023 |
| Senior Director of Research | Oriol Vinyals | Former Google Brain researcher, technical lead on Gemini |
| Co-founder (departed) | Mustafa Suleyman | Left DeepMind 2019; left Google 2022; co-founded Inflection AI 2022; CEO of Microsoft AI from March 2024 |
Additional senior figures include David Silver, who led the AlphaGo, AlphaZero, and MuZero projects; John Jumper, director of the AlphaFold team and co-recipient of the 2024 Nobel Prize in Chemistry; Karen Simonyan, who left in 2022 to co-found Inflection AI and later joined Microsoft AI alongside Suleyman; and Josh Woodward, formerly head of Google Labs, who took over the lead role for Gemini consumer product development in April 2025.[24][29]
Google DeepMind is wholly owned by Alphabet, Google's parent company, and operates as the central AI research and development unit within Alphabet following the 2023 merger. It is not a separately traded company and has no outside shareholders. Estimates of total headcount vary by source and by whether one counts the entire merged division or only the legacy DeepMind Technologies subsidiary. Public filings by DeepMind Technologies Limited at UK Companies House and headcount trackers indicate that DeepMind Technologies grew from roughly 2,500 employees at the time of the merger to approximately 3,200 by the end of 2024 and around 4,500 by late 2025. Reporting in 2025 estimated that the broader merged Google DeepMind division employed in the range of 5,000 to 6,000 staff worldwide, including researchers, engineers, product managers, and operations personnel.[30][31]
The global headquarters is at 6 Pancras Square in King's Cross, central London, in a Google-occupied building that also houses other Google teams. Additional substantial DeepMind sites operate at Google offices in Mountain View and New York in the United States, in Zurich and Paris in Europe, in Tokyo, Toronto, Tel Aviv, Bangalore, and Seattle. Many specialist teams (Gemini training, AlphaFold, robotics, weather, materials) are split across multiple sites, with the Gemini effort in particular spanning London, Mountain View, and Zurich. In 2025, Google announced a new building called Platform 37 alongside 6 Pancras Square that would house additional Google DeepMind staff and a research-focused "AI Exchange."[32]
Financially, DeepMind Technologies operated at significant losses throughout the 2010s and early 2020s, supported by intercompany funding from Alphabet. UK filings revealed losses of about $649 million for 2019, with Alphabet waiving roughly $1.5 billion in intercompany debt in the same period. After the 2023 merger, Google ceased reporting DeepMind as a stand-alone segment of Alphabet, and Alphabet now categorizes AI investment within its Google Services and Google Cloud businesses.[33]
The following table summarizes selected projects and events from DeepMind and Google DeepMind, with the year referring to the public announcement, paper publication, or completion of the milestone in question.[3][4]
| Year | Project / event | Description |
|---|---|---|
| 2010 | Founding | DeepMind Technologies founded in London by Demis Hassabis, Shane Legg, and Mustafa Suleyman |
| 2013 | Atari DQN | NIPS workshop paper on deep Q-network playing Atari games from raw pixels |
| 2014 | Google acquisition | Acquired on January 26, 2014 for a reported figure in the $400-650 million range |
| 2015 | Nature DQN | "Human-level control through deep reinforcement learning" published in Nature |
| 2015 | Fan Hui match | AlphaGo defeats European Go champion Fan Hui 5-0 in October 2015 |
| 2016 | AlphaGo vs Lee Sedol | AlphaGo defeats world champion Lee Sedol 4-1 in Seoul in March 2016 |
| 2016 | Data center cooling | DeepMind reduces Google data center cooling energy use by 40% |
| 2016 | WaveNet | Generative model for raw audio waveforms; later adopted in Google Assistant |
| 2017 | AlphaGo Zero | Self-play-only Go agent surpasses earlier AlphaGo versions |
| 2017 | DeepMind Ethics & Society | Public ethics research unit launched in October 2017 |
| 2018 | AlphaZero | Single algorithm masters chess, shogi, and Go from self-play |
| 2018 | AlphaFold 1 | Wins the CASP13 protein structure prediction competition |
| 2019 | AlphaStar | Reaches Grandmaster level in StarCraft II in all three races |
| 2019 | MuZero | Plans with a learned model across Atari, Go, chess, and shogi |
| 2019 | Suleyman leave | Mustafa Suleyman placed on leave; later departs DeepMind |
| 2020 | AlphaFold 2 | Solves CASP14 with median GDT_TS of 92.4; described as solving the protein folding grand challenge |
| 2021 | AlphaFold release | Source code released; AlphaFold Protein Structure Database launched with EMBL-EBI |
| 2022 | AlphaCode | First AI to reach median human level on Codeforces programming contests |
| 2022 | Chinchilla | Compute-optimal scaling laws paper redefines language model training |
| 2022 | Gato | Single 1.2B-parameter generalist agent across 604 tasks |
| 2022 | AlphaTensor | Discovers faster matrix multiplication algorithms (Nature, October 2022) |
| 2022 | DeepMind 200M | AlphaFold database expanded to ~200 million predicted protein structures |
| 2023 | Merger | DeepMind and Google Brain merge into Google DeepMind on April 20, 2023 |
| 2023 | SynthID | Watermarking system for AI-generated images launches in August 2023 |
| 2023 | GraphCast | ML weather forecasting beats ECMWF on 90% of evaluation targets |
| 2023 | GNoME | 2.2 million predicted stable crystal structures, 381,000 newly stable |
| 2023 | Lyria | High-quality music generation model announced with YouTube |
| 2023 | Gemini 1.0 | Natively multimodal model family launched on December 6, 2023 |
| 2023 | FunSearch | LLM-based mathematical discovery, advances on cap-set problem (Nature) |
| 2024 | Gemini 1.5 | Mixture-of-experts with 1M-2M token context windows |
| 2024 | Project Astra | Universal AI assistant prototype unveiled at Google I/O |
| 2024 | Veo | First-generation text-to-video model |
| 2024 | Imagen 3 | Highest-quality Google text-to-image model at the time |
| 2024 | AlphaFold 3 | Diffusion-based, predicts proteins, DNA, RNA, ions, ligands |
| 2024 | AlphaProof / AlphaGeometry 2 | Silver-medal score (28/42) at IMO 2024 |
| 2024 | AlphaProteo | De novo protein binder design system |
| 2024 | Nobel Prize | Hassabis and Jumper share Nobel Prize in Chemistry |
| 2024 | Project Mariner | Browser-control AI agent unveiled in December 2024 |
| 2024 | Genie 2 | Foundation world model generating playable 3D environments |
| 2024 | Gemini 2.0 Flash | Agentic, natively-multimodal-output model launched in December |
| 2025 | Gemini 2.5 / Deep Think | Reasoning model series, gold-medal IMO performance later in year |
| 2025 | Veo 3 | Text-to-video with synchronized native audio and dialogue |
| 2025 | Genie 3 | Higher-resolution, longer-horizon world model |
| 2025 | Gemini Robotics | Vision-language-action models for physical robots |
| 2025 | SIMA 2 | Gemini-powered generalist agent for 3D virtual worlds |
| 2025 | Gemini 3 Pro | Flagship Gemini 3 family launched on November 18, 2025 |
| 2026 | Lyria 3 | Music generation model launched in February 2026 |
Reinforcement learning has been DeepMind's signature research area since its founding. The 2013 NIPS DQN paper, with its single network learning Atari games from raw pixels, demonstrated the viability of deep RL on perceptual control tasks. The 2015 Nature paper extended this to all 49 evaluated Atari 2600 games, establishing the benchmark on which most subsequent deep RL work was measured. AlphaGo combined Monte Carlo tree search (MCTS) with deep neural networks trained via supervised learning on human games and refined via self-play RL. AlphaGo Zero and AlphaZero stripped away the human game data, demonstrating that pure self-play, given only the rules and a sufficient compute budget, can produce superhuman play.[14][16]
MuZero went one step further, learning a model of environment dynamics in a latent space rather than relying on access to the rules, and unifying model-based and model-free RL. Alongside these high-profile games-of-strategy results, DeepMind has produced foundational work on distributional RL (C51, QR-DQN), exploration (Bootstrapped DQN, NoisyNet), off-policy learning (Retrace, V-trace, IMPALA), population-based training, and large-scale multi-agent RL.[17][18]
In parallel, RL techniques have been applied to real-world optimization problems within Alphabet, including the 2016 data center cooling work, chip floorplanning (with the 2021 Nature paper on RL-based chip placement), Google Maps ETA prediction, and YouTube recommendation. Inside the Gemini lab, RL has returned to a central role through reinforcement learning from human feedback (RLHF), reinforcement learning from AI feedback (RLAIF), and the search-based reasoning techniques that underpin Gemini Deep Think mode.[34]
AlphaFold is widely regarded as the most impactful application of AI to science. AlphaFold 1, deployed in CASP13 in 2018, used convolutional neural networks to predict pairwise residue distance distributions, which were then converted into structures via gradient-based optimization. AlphaFold 2, by contrast, used the Evoformer attention module operating jointly on a multiple sequence alignment and pairwise residue features, with an end-to-end structure module producing 3D coordinates. The 2020 CASP14 and 2021 Nature results were a discontinuous improvement over previous methods.[19]
AlphaFold 3, announced on May 8, 2024 in a Nature paper by Josh Abramson, John Jumper, and colleagues, replaced the structure module with a diffusion-based generative head and extended the predictive scope from single-protein structures to a full unified model of protein-protein, protein-nucleic-acid, protein-ligand, protein-ion, and protein-modification interactions. Reported gains include 76 percent of protein-ligand interactions predicted accurately (versus 52 percent for the previous best dedicated software), 65 percent of DNA interactions (versus 28 percent), and roughly double AlphaFold 2's accuracy on protein-protein interactions. AlphaFold 3 was made available to non-commercial researchers via the AlphaFold Server interface, while Isomorphic Labs, a 2021 spinout from DeepMind under Hassabis, retains commercial rights for drug discovery.[35]
DeepMind's broader "AI for biology" agenda includes AlphaProteo, announced in September 2024, which is a system for de novo design of high-affinity protein binders. AlphaProteo achieved 3- to 300-fold better binding affinities than the best previous design methods on seven target proteins, including the SARS-CoV-2 spike receptor-binding domain, VEGF-A, IL-7Ra, PD-L1, TrkA, IL-17A, and the Epstein-Barr virus protein BHRF1. In collaboration with the Francis Crick Institute, AlphaProteo binders for the SARS-CoV-2 spike were tested and shown to neutralize viral entry into human cells. The lab has also published on AlphaMissense (predicting pathogenicity of missense variants), genomic sequence prediction, and antibody design.[36]
GraphCast, published in Science in November 2023, applies graph neural networks to medium-range global weather forecasting. Trained on 39 years of ERA5 reanalysis data from the European Centre for Medium-Range Weather Forecasts (ECMWF), GraphCast produces 10-day global forecasts at 0.25-degree resolution (around 28 km at the equator) in under one minute on a single Google TPU v4. In ECMWF-style evaluation it outperformed the operational High-Resolution Forecast (HRES) on more than 90 percent of 1,380 evaluation targets covering temperature, wind, geopotential height, and humidity at multiple atmospheric levels.[37]
GenCast, announced in late 2024, extended the approach to probabilistic ensemble forecasting using diffusion models, providing calibrated uncertainty estimates that beat the ECMWF ensemble system on a substantial fraction of metrics. In 2025, DeepMind announced a specialized cyclone prediction model that uses stochastic neural networks trained on 45 years of global weather and tropical cyclone data, producing probabilistic forecasts of cyclone formation, track, intensity, and structure up to 15 days in advance. Components of GraphCast and GenCast have been integrated into Google Search, Google Maps, and weather products.[38]
FunSearch, introduced in a December 2023 Nature paper by Bernardino Romera-Paredes and colleagues, pairs an LLM with an automated program-correctness evaluator inside an evolutionary search loop. The system found the largest known cap sets in dimensions 8 through 12 (the most significant progress on the cap set problem in over two decades) and improved heuristics for the online bin-packing problem that beat human-designed methods. FunSearch was the first published demonstration of an LLM producing genuinely novel verified mathematical results.[39]
In July 2024, DeepMind announced that AlphaProof and AlphaGeometry 2, a paired system, had together solved four of the six problems at the 2024 International Mathematical Olympiad (IMO), scoring 28 out of 42 points. This is the same score that earned a silver medal in the human contest. AlphaProof handled three problems (two algebra, one number theory) by translating them into the Lean proof assistant and applying RL-driven search; AlphaGeometry 2 (a successor to AlphaGeometry, announced January 2024) solved the geometry problem using a neuro-symbolic hybrid in which a Gemini-trained language model proposes auxiliary constructions and a symbolic engine completes the proofs. A subsequent IMO-style evaluation in 2025 with a more advanced Gemini Deep Think model achieved gold-medal-level performance, solving five of six problems.[40]
Other scientific results include AlphaTensor, which in October 2022 used RL to discover faster matrix multiplication algorithms in modular arithmetic, and AlphaDev, which in June 2023 found faster sorting and hashing algorithms now incorporated into the LLVM standard library. The GNoME (Graph Networks for Materials Exploration) project, published in Nature in November 2023, used graph neural networks and active learning to predict 2.2 million stable inorganic crystal structures, of which 381,000 were both novel and predicted to be stable. The Materials Project at Lawrence Berkeley National Laboratory has incorporated GNoME predictions, and external groups have already synthesized hundreds of the predicted compounds in the laboratory.[41][42]
The Gemini series is Google DeepMind's flagship family of large language models. Gemini was announced and partly previewed by Pichai at Google I/O in May 2023 and formally launched on December 6, 2023 at a virtual press conference, with three sizes (Ultra, Pro, Nano) targeted at different deployment regimes. Gemini was designed from the start to be natively multimodal, with text, image, audio, and video tokens trained jointly rather than appended to a text base, and was the first major flagship LLM to do so.[26][43]
The Gemini 1.5 line, announced in February 2024, introduced a sparse mixture-of-experts (MoE) architecture and very long context windows (up to 1 million tokens, later 2 million in some configurations). Gemini 2.0 Flash, announced in December 2024, was DeepMind's first "agentic" model, capable of producing native multimodal output (text, image, audio) and supporting complex tool use. Gemini 2.5, announced in March 2025, introduced an explicit reasoning step, the "thinking" mode in which the model is allowed to produce extended internal reasoning chains before committing to an answer. Gemini 2.5 Deep Think, previewed at Google I/O on May 20, 2025 and made generally available on August 1, 2025, took this further with parallel-search reasoning and posted very high scores on USAMO and competitive coding benchmarks. Gemini 3 Pro, launched on November 18, 2025, became the first publicly available model to surpass an Elo of 1500 on the LMArena leaderboard.[44][45][66]
| Version | Released | Notable characteristics |
|---|---|---|
| Gemini 1.0 (Ultra, Pro, Nano) | December 6, 2023 | Natively multimodal across text, image, audio, video, and code |
| Gemini 1.5 Pro | February 2024 | MoE architecture, 1M token context (later expanded to 2M for select users) |
| Gemini 1.5 Flash | May 2024 | Cost-efficient, lower-latency variant for developers |
| Gemini 2.0 Flash | December 2024 | Native multimodal output and tool-use; agentic features |
| Gemini 2.5 Pro | March 2025 | First Gemini designated a reasoning ("thinking") model |
| Gemini 2.5 Flash / Flash-Lite | May-June 2025 | Faster, cheaper reasoning variants |
| Gemini 2.5 Deep Think | August 2025 | Parallel-search reasoning, gold-medal IMO performance |
| Gemini 3 Pro | November 18, 2025 | Flagship of Gemini 3 family; first model above 1500 LMArena Elo |
| Gemini 3 Flash | December 2025 | Fast variant of Gemini 3 |
| Gemini 3.1 Pro | Early 2026 | Successor iteration, latest available model |
Gemini models power Google Search AI Mode and AI Overviews, the standalone Gemini chatbot app, Google Workspace integrations such as Gemini for Docs and Gmail, the Android operating system through on-device Gemini Nano, the Pixel device line, and developer products via the Gemini API and Vertex AI. On the research side, the Gemini program incorporates lineage from Google Brain's PaLM, PaLM 2, T5, and Pathways efforts, as well as DeepMind's Chinchilla, Gopher, and Sparrow projects.[27]
DeepMind has produced a steady stream of generative models across modalities since 2016. WaveNet, introduced in the September 2016 paper by Aaron van den Oord and colleagues, was a deep autoregressive convolutional model for raw audio waveforms; it powered substantially more natural text-to-speech in Google Assistant starting in 2017 and remains a milestone in neural audio synthesis. Tacotron and successors built on similar principles for end-to-end speech synthesis.[46]
Imagen, originally introduced by Google Research in May 2022 as a text-to-image diffusion model, became part of the unified Google DeepMind generative product lineup after the merger. Imagen 2 was launched in December 2023, Imagen 3 was announced at Google I/O in May 2024 and made broadly available in August 2024, and subsequent versions ship with Veo and Gemini Image generation features. Lyria, announced in November 2023 with YouTube, is DeepMind's high-quality music generation model; the original Lyria powered the Dream Track experiment for YouTube Shorts and the Music AI Sandbox for professional musicians. Lyria 2 followed in 2024, and Lyria 3 launched in February 2026 inside the Gemini app.[47][48]
Veo, DeepMind's flagship text-to-video model, was first announced in May 2024 with the ability to generate 1080p clips of over a minute in length. Veo 2, in December 2024, added 4K resolution and improved physics. Veo 3, launched in May 2025, introduced joint audio-visual generation: it produces synchronized dialogue, sound effects, and ambient noise alongside the video, processing audio at 48 kHz stereo and rendering at up to 4K, 24 frames per second, in 16:9 or 9:16 aspect ratios. Veo 3.1 followed in October 2025 with substantial improvements to prompt adherence and audio-visual alignment. Veo is integrated into the Gemini App, the Vertex AI platform, the Vids product, and YouTube Shorts via Dream Screen.[49]
Gato, described in the May 2022 paper "A Generalist Agent" by Scott Reed and colleagues, was a single 1.2-billion-parameter transformer trained on 604 distinct tasks across modalities, including Atari games, image captioning, dialogue, and real robot block-stacking. Gato was an early demonstration that a single model with a single set of weights could output text tokens, joint torques, button presses, and other action tokens depending on context. SIMA, introduced in March 2024 (Scalable Instructable Multiworld Agent), is a generalist agent trained to follow natural-language instructions across nine commercial 3D video games (including No Man's Sky and Teardown) using only screen pixels and standard keyboard-and-mouse output. SIMA 2, announced in November 2025, integrates Gemini reasoning with the SIMA action interface, allowing the agent to converse with users about its goals and improve its own behavior.[50][51]
Project Astra, unveiled at Google I/O in May 2024, is DeepMind's research prototype for a "universal AI assistant" that takes in continuous audio and video input via a phone camera or smart glasses and answers questions about the user's environment in real time. It was iterated through 2025 and has informed the Gemini Live multimodal interaction features in the Gemini app. Project Mariner, announced in December 2024 and described in the dedicated Project Mariner entry, is DeepMind's browser-using agent: built on Gemini 2.0, it controls Chrome via a research extension, achieving 83.5 percent on the WebVoyager benchmark and supporting up to 10 simultaneous tasks.[52]
Genie, presented in the February 2024 paper "Genie: Generative Interactive Environments" by Jake Bruce and colleagues, was the first foundation "world model" trained on internet videos in an unsupervised fashion. The 11-billion-parameter model could generate playable 2D environments controllable on a frame-by-frame basis from a single image prompt. Genie 2, announced in December 2024, generated longer-horizon, action-controllable 3D environments. Genie 3, announced in August 2025, achieved higher resolution and several minutes of visual consistency. Genie 3 was made available to Google AI Ultra subscribers in early 2026.[53]
In March 2025, Google DeepMind launched Gemini Robotics and Gemini Robotics-ER (Embodied Reasoning), a pair of vision-language-action models that extend Gemini's understanding into the physical world. Gemini Robotics was designed to control robot manipulators given natural-language task descriptions and visual input, while Gemini Robotics-ER focused on physical reasoning and the planning of multi-step tasks. Gemini Robotics 1.5, announced in September 2025, improved both dexterity and reliability across two-arm manipulation benchmarks. The robotics work draws on DeepMind's earlier RT-1, RT-2, and RoboCat lines and on the Robotics Transformer family from Google. The detailed Gemini Robotics article covers this program in greater depth.[54]
DeepMind has maintained a substantial AI safety research program since well before the merger. Public outputs include the 2017 "Specifying AI safety problems in simple environments" paper, the AI Safety Gridworlds benchmark, the Sparrow research dialogue agent (2022), and the SynthID watermarking system. SynthID, first released in August 2023 for images generated by Imagen, embeds an imperceptible watermark into pixels at synthesis time. By May 2024 SynthID had been extended to text outputs from Gemini, audio outputs from Lyria, and video outputs from Veo. SynthID Text was published in Nature in October 2024 ("Scalable watermarking for identifying large language model outputs"), and the text decoder was released open-source via Google's Responsible GenAI Toolkit and on Hugging Face.[55][56]
In May 2024, Google DeepMind introduced its Frontier Safety Framework, a set of protocols for evaluating frontier models against "Critical Capability Levels" (CCLs), thresholds at which a model's capabilities cross into territory requiring additional risk mitigation. The framework follows the same general spirit as Anthropic's Responsible Scaling Policy and OpenAI's Preparedness Framework. Version 2 was published in 2025 and Version 3 in 2026, expanding coverage to include misuse CCLs (cyber capability, biological and chemical risk, manipulation), machine learning R&D CCLs (whether a model can substantially accelerate AI research itself), and misalignment CCLs. The Frontier Safety Framework is one of the artifacts that DeepMind cites in its public commitments to AI Safety Summit pledges.[57]
A dedicated AGI Safety and Alignment team at Google DeepMind, led for several years by Rohin Shah and David Krueger, has produced influential work on scalable oversight, debate, model evaluations for dangerous capabilities, scheming behavior detection, and interpretability. The lab also runs internal red-teaming, an AI Principles review committee, and external collaborations with the UK AI Safety Institute and the US AI Safety Institute on pre-deployment evaluations of frontier models.[57]
Beyond the headline projects, DeepMind has contributed to many other areas of computational science. AlphaTensor (October 2022, Nature) used RL to discover faster matrix multiplication algorithms, including the first improvement on Strassen's two-level 4x4 algorithm in modular arithmetic in over 50 years. AlphaDev (June 2023, Nature) discovered faster sorting and hashing routines that were merged into the LLVM libc++ standard library, with measurable speedups across millions of devices. The lab also reported, in February 2024, that DeepMind controllers had stabilized plasma confinement in the Tokamak a Configuration Variable (TCV) at EPFL, work conducted with the Swiss Plasma Center.[42][58]
In the medical and clinical sphere, DeepMind Health was an internal unit until late 2018, when its operational health products (most notably the Streams patient-monitoring app developed with the UK National Health Service) were transferred to Google Health, in a transition that drew controversy due to questions about the original NHS data-sharing agreement. The current lab has continued health-relevant research, including AlphaMissense for variant pathogenicity prediction, work on diabetic retinopathy with Verily, and large-scale predictions over the human proteome.[3]
Google DeepMind sits within Alphabet alongside several related organizations. Google Research retains responsibility for parts of computer science, hardware, and applied research, while Google DeepMind concentrates on the core frontier-model and applied-AI research stack. Jeff Dean serves as chief scientist of both organizations, providing a coordination link. Google Cloud's AI products are built on top of Gemini and other DeepMind models, exposed through the Vertex AI platform.[24]
Isomorphic Labs, founded in November 2021 by Hassabis as an Alphabet subsidiary, applies AlphaFold-derived methods to drug discovery. Isomorphic Labs is a separate company from Google DeepMind but shares some research and a commercial license to AlphaFold 3 for drug-design applications. In 2025 Isomorphic Labs raised $600 million in external investment from Thrive Capital and others, and announced its first internal-pipeline drug candidates entering preclinical study, with the goal of beginning human trials.[59]
Waymo, Verily, Calico, and Google X (the moonshot factory) are other Alphabet entities that occasionally collaborate with Google DeepMind on AI components, but they remain operationally independent.
In the second half of the 2020s, Google DeepMind has become one of three frontier-model labs that dominate public discussion, alongside OpenAI (founded 2015, headquartered in San Francisco) and Anthropic (founded 2021, also in San Francisco, by former OpenAI researchers including Dario Amodei). Google DeepMind is the largest of the three by headcount, by capital backing (through Alphabet), and by access to in-house compute (Google's TPU fleet). It is also distinctive in two other respects: it operates the broadest portfolio of applied scientific projects (AlphaFold, GraphCast, GNoME, AlphaTensor, AlphaProof, AlphaProteo) of any major AI lab, and it is the only frontier lab whose CEO is a Nobel laureate.[60]
Unlike OpenAI and Anthropic, both of which are largely API-and-product companies, Google DeepMind also remains highly active in pure scientific publication. The lab regularly contributes papers to Nature, Science, NeurIPS, ICML, ICLR, and major domain conferences in biology, chemistry, and weather science. As of 2026, the lab estimates that it has co-authored more than 5,000 peer-reviewed publications since founding.[3]
As of mid-2026, Google DeepMind operates as the central AI research and development unit within Alphabet, with Demis Hassabis reporting to Sundar Pichai and, by his own account in a January 2026 CNBC interview, communicating with Pichai daily on AI strategy. Hassabis has stated publicly that AGI may arrive within roughly the next decade and possibly sooner, and that the lab's research priorities are organized around that thesis. In the company's own framing, "AI, and ultimately artificial general intelligence (AGI), has the potential to drive one of the greatest transformations in history."[63] The organization was named to TIME's 100 Most Influential Companies list for 2025.[28][61]
The lab continues to advance across multiple frontiers in parallel: scaling the Gemini model family (with Gemini 3.1 Pro the most recent flagship as of early 2026); pushing scientific AI applications through AlphaFold 3, AlphaProteo, GNoME, and the cyclone-prediction work; developing world models through Genie; building safety infrastructure through SynthID and the Frontier Safety Framework; advancing robotic intelligence through Gemini Robotics; and prosecuting AI agent research through Project Astra, Project Mariner, and SIMA 2. The 2023 merger has given the combined entity end-to-end ownership of training data, compute, model research, evaluation, and product, while preserving DeepMind's research culture and its practice of publishing in top scientific journals.
Josh Woodward, previously head of Google Labs, was moved into the lead role for Gemini consumer product development in April 2025, reflecting the strategic importance of Gemini to Alphabet's overall business. The lab's policy and government engagement is led by Lila Ibrahim, now Chief AI Readiness Officer, who has represented Google DeepMind at the UK and Korea AI Safety Summits and at multiple G7 and OECD AI working groups.[62]
The following Google DeepMind topics are now covered on their own pages: