UK BUSINESS AI UNIVERSE: Labs: General-purpose AI models

General-purpose AI models

3 labs · 75 models and products Theme: Companies · Generated from this page's data. Click a name to open its panel on the Labs page. Records resting on one source are tagged "one source".

General-purpose AI models, also called foundation models, are large systems trained on very large amounts of text, images, sound and code, which can then be adapted to many tasks, from chat to coding. On this tile Google DeepMind builds the Gemini family, the open Gemma models and the Genie "world models", which generate playable environments; the tile lists 63 models and products from DeepMind, Locai, Recursive and Stability AI (per the site). The UK's competition regulator, the CMA, has reviewed how these models affect competition and consumers since 2023.

Context

The Competition and Markets Authority published its first report on foundation models and proposed principles for competitive AI markets. 18 September 2023 gov.uk/government/news/proposed-principles-to-guFMs are artificial intelligence (AI) systems with broad capabilities that can be adapted to a range of different
The CMA followed with an update report on foundation models. 11 April 2024 gov.uk/cma-cases/ai-foundation-models-initial-reUpdate report published
Google DeepMind's Gemini family was launched in three sizes. 19 December 2023 arxiv.org/abs/2312.11805The Gemini family consists of Ultra, Pro, and Nano sizes
Gemma, Google DeepMind's open-weight models, beat open models of similar size on 11 of 18 text tasks, its makers report. 13 March 2024 arxiv.org/abs/2403.08295Gemma outperforms similarly sized open models on 11 out of 18 text-based tasks
DSIT's AI sector study found the highest UK AI investment values in 2025 went to companies developing and training AI. 29 October 2026 gov.uk/government/publications/artificial-intellHowever, the largest investment values in 2025 were directed towards AI development and training…
The government's AI Opportunities Action Plan described Britain as third in the world as an AI market. 13 January 2025 gov.uk/government/publications/ai-opportunities-Today, Britain is the third largest AI market in the world.

Key papers

Gemini: A Family of Highly Capable Multimodal Models arXiv preprint · 19 December 2023
Gemma: Open Models Based on Gemini Research and Technology arXiv preprint · 13 March 2024
Genie: Generative Interactive Environments arXiv preprint · 23 February 2024

Labs Round 8 (31 Oct 2026): introduction and facts as the researchers wrote them; "(per the site)" marks a sentence resting on this site's own description of that lab.

Record	What it is	Place	Links	Confidence
AeneasChips, models and products	An AI model that helps historians date, place and restore ancient Latin inscriptions. released (free tool; open code and data)Key facts Nature paper 23 Jul 2025; free interactive version at predictingthepast.com; code and dataset on GitHub. Figures (23 Jul 2025): 73% top-20 restoration accuracy for gaps up to ten characters; 72% province attribution across 62 provinces; dataset of over 176,000 Latin inscriptions. Predecessor: Ithaca (ancient Greek, March 2022).Made by Google DeepMind · 2025-07-23	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/aeneas-transforms-how-histo · deepmind.google/blog/page/12	one source
AlphaCodeChips, models and products	A system that writes computer programs for programming contests by generating many candidates and filtering them. released (paper and dataset)Key facts First blog 2 Feb 2022; Science cover paper 8 Dec 2022. Access: paper; problems-and-solutions dataset on GitHub; solutions viewable at alphacode.deepmind.com. Figures (8 Dec 2022): estimated rank within top 54% of contest participants; "about the level of the median competitor".Made by Google DeepMind · 2022-02-02Versions AlphaCode → AlphaCode 2Read the paper	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/competitive-programming-wit · deepmind.google/blog/page/10/	one source
AlphaEarth FoundationsChips, models and products	An AI model that combines satellite and other Earth data into one map-like representation for tracking land and environment. released (Satellite Embedding dataset in Google Earth Engine)Key facts Announced 30 Jul 2025. Figures (30 Jul 2025): 10x10 metre squares; 24% lower error rate than other approaches tested; over 1.4 trillion embedding footprints per year; used by more than 50 organisations. Access: annual embeddings in Earth Engine.Made by Google DeepMind · 2025-07-30	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/alphaearth-foundations-help · deepmind.google/models/	lab's own pages
AlphaFold (first version, CASP13)Chips, models and products	The first AlphaFold, which predicts a protein's 3D shape from its sequence; it ranked first at the CASP13 benchmark in 2018. released (paper)Key facts Lab timeline: placed first at CASP13 (2 Dec 2018); methods later published in Nature. Succeeded by AlphaFold2 (Nov 2020) and AlphaFold 3 (May 2024).Made by Google DeepMind · 2018-12-02	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/science/alphafold/	one source
AlphaFold 3 and AlphaFold ServerChips, models and products	The third AlphaFold, which predicts the structures and interactions of proteins and other molecules, offered through a free web server. released (free server; code and weights for academic use)Key facts Launched 8 May 2024 (lab timeline). Access: AlphaFold Server free for non-commercial research; model code and weights for academic use. Usage (lab AlphaFold page, figure undated, read 30 Oct 2026): "over 3 million researchers from over 190 countries" use AlphaFold. Nobel Prize in Chemistry shared for AlphaFold, 9 Oct 2024.Made by Google DeepMind · 2024-05-08Read the paper Key facts and news (1) Models released AlphaFold 3 model code and weights offered for academic use (download via GitHub; release date not stated on the page); AlphaFold Server gives free AF3 predictions for non-commercial research from 8 May 2024 one source deepmind.google/science/alphafold/ Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked.	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/science/alphafold/ · deepmind.google/blog/page/8/	lab's own pages
AlphaFold2Chips, models and products	The second AlphaFold, which predicts protein shapes close to lab-measured accuracy; organisers called it a solution to protein folding. released (open code)Key facts Announced 30 Nov 2020 after CASP14. Figures (30 Nov 2020): median 92.4 GDT across all targets; average error about 1.6 Angstroms. Methods published in Nature with open-source code 15 Jul 2021 (lab timeline).Made by Google DeepMind · 2020-11-30Read the paper	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/alphafold-a-solution-to-a-5 · deepmind.google/science/alphafold/	lab's own pages
AlphaGo ZeroChips, models and products	A Go-playing system that taught itself the game from random moves by playing against itself, with no human game records. released (paper)Key facts Published 18 Oct 2017 in Nature. Access: paper only; game records downloadable. Figures (18 Oct 2017): beat the champion-defeating AlphaGo 100 games to 0; overtook AlphaGo Master after 40 days of self-training; Elo chart puts it over 5,000. Version history (as published): AlphaGo Fan → AlphaGo Lee → AlphaGo Master → AlphaGo Zero.Made by Google DeepMind · 2017-10-18Versions AlphaGo Fan → AlphaGo Lee → AlphaGo Master → AlphaGo ZeroRead the paper	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/alphago-zero-starting-from- · deepmind.google/blog/page/17/	one source
AlphaQubitChips, models and products	An AI system aimed at quantum error correction (per its post address). research releaseKey facts November 2024 (index; post on blog.google not opened).Made by Google DeepMind · 2024-11	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/8/	one source
AlphaStarChips, models and products	An AI program that plays the full real-time strategy game StarCraft II and beat top professional players. released (research result)Key facts Announced 24 Jan 2019. Access: no product; replays downloadable; journal write-up in preparation at announcement. Figures (24 Jan 2019): beat professional players TLO and MaNa 5–0 each; trained for 14 days using 16 TPUs per agent; average around 280 actions per minute.Made by Google DeepMind · 2019-01-24Read the paper	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/alphastar-mastering-the-rea · deepmind.google/blog/page/15/	one source
AlphaZeroChips, models and products	One learning system that teaches itself chess, shogi and Go from the rules alone, by playing against itself. released (paper)Key facts Full evaluation published 6 Dec 2018 in Science (open-access preprint). Access: paper only; game files downloadable. Figures (6 Dec 2018): chess v Stockfish won 155 and lost 6 of 1,000 games; shogi v Elmo won 91.2% of games; Go v AlphaGo Zero won 61% of games. Predecessor named: AlphaGo Zero.Made by Google DeepMind · 2018-12-06Read the paper	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/alphazero-shedding-new-ligh · deepmind.google/blog/page/16	one source
ChinchillaChips, models and products	A 70-billion-parameter language model that showed training on more data can beat simply adding parameters. released (paper)Key facts Blog 12 Apr 2022; arXiv 2203.15556. Figures (12 Apr 2022): 70 billion parameters trained on 1.3 trillion tokens; compared with Gopher's 280 billion parameters. Access: paper only.Made by Google DeepMind · 2022-04-12Read the paper	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/an-empirical-analysis-of-co	one source
CodeGemmaChips, models and products	Small open Gemma models trained to complete, write and chat about computer code. released (open weights)Key facts Initial release 9 Apr 2024; v1.1 released 3 May 2024; variants: 7B pretrained, 7B instruction-tuned, 2B pretrained (stated "up to 2x faster code completion"); trained on 500 billion tokens; access: open weightsMade by Google DeepMind · 2024-04-09Versions Gemma → Gemma 2 (and CodeGemma, RecurrentGemma, PaliGemma → PaliGemma 2, ShieldGemma → ShieldGemma 2, DataGemma)	London	ai.google.dev/gemma/docs/codegemma · ai.google.dev/gemma/docs/releases	lab's own pages
D4RTChips, models and products	A model for reconstructing and tracking scenes in four dimensions (3D plus time), claimed to be much faster than earlier methods. researchKey facts Published 16 Jan 2026; claimed up to 300x faster than prior methods.Made by Google DeepMind · 2026-01-16	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml	one source
DataGemma (RIG and RAG)Chips, models and products	Gemma 2 models tuned to answer plain-language questions with public statistics drawn from Google's Data Commons. released (open weights; research tool)Key facts Released 12 Sep 2024; two approaches: Retrieval-Interleaved Generation (RIG) and Retrieval-Augmented Generation (RAG, which feeds Gemini 1.5 Pro); access: open weightsMade by Google DeepMind · 2024-09-12Versions Gemma → Gemma 2 (and CodeGemma, RecurrentGemma, PaliGemma → PaliGemma 2, ShieldGemma → ShieldGemma 2, DataGemma)	London	ai.google.dev/gemma/docs/datagemma · ai.google.dev/gemma/docs/releases	one source
Decoupled DiLoCoChips, models and products	A Google DeepMind method for resilient training of AI across distributed computers. researchKey facts Post dated 22 Apr 2026 (title only opened).Made by Google DeepMind · 2026-04-22	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml	one source
DiffusionGemmaChips, models and products	An open Gemma model that generates text by diffusion, which the lab says is four times faster. released (open weights)Key facts Theme: General-purpose AI models. Released 10 Jun 2026; model card same date.Made by Google DeepMind · 2026-06-10	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml · deepmind.google/models/model-cards/	lab's own pages
DolphinGemmaChips, models and products	A Gemma-based model used to help decode dolphin communication. research releaseKey facts Theme: General-purpose AI models. April 2025 (index title); listed on the Gemma page.Made by Google DeepMind · 2025-04	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/7/	one source
EmbeddingGemma (and EmbeddingGemma 2)Chips, models and products	Small open Gemma models that turn text (and, in version 2, other media) into numbers for search and retrieval. released (open weights)Key facts Theme: General-purpose AI models. EmbeddingGemma September 2025 (card 25 Sep 2025); EmbeddingGemma 2 released 6 Oct 2026 as an open, lightweight multimodal embedding model (card same date).Made by Google DeepMind · 2025-09	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml · deepmind.google/models/model-cards/	lab's own pages
FlamingoChips, models and products	A single model that reads images, video and text together and learns new tasks from a few examples. released (paper)Key facts Preprint announced 28 Apr 2022. Figures (28 Apr 2022): 80-billion-parameter visual language model; studied on 16 tasks. Access: paper only.Made by Google DeepMind · 2022-04-28	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/tackling-multiple-tasks-wit · deepmind.google/blog/page/12	one source
FunctionGemma and TranslateGemmaChips, models and products	Specialised open Gemma models: one for calling software functions, one for translation. released (open weights)Key facts Theme: General-purpose AI models. FunctionGemma December 2025 (Gemma page; card 14 Jan 2026); TranslateGemma January 2026 (Gemma page).Made by Google DeepMind · 2025-12	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/models/gemma/ · deepmind.google/models/model-cards/	lab's own pages
Gemini 1.0Chips, models and products	Google DeepMind's first Gemini model family, described at launch as its largest and most capable AI model. released (superseded)Key facts Announced December 2023 (blog index). Technical report linked from the model-cards page (no date shown there). Launch post on blog.google could not be opened in this session, so sizes and benchmark figures are not recorded.Made by Google DeepMind · 2023-12	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/9/ · deepmind.google/models/model-cards/ · arxiv.org/abs/2312.11805	lab's own pages
Gemini 1.5Chips, models and products	The second Gemini generation; a May 2024 update added a faster model and a longer context window. released (superseded)Key facts Gemini 1.5 announced February 2024; "Gemini breaks new ground: a faster model, longer context and AI agents" May 2024; updated production models and reduced 1.5 Pro pricing September 2024 (developers blog). Launch posts on blog.google could not be opened.Made by Google DeepMind · 2024-02Read the paper	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/9/ · deepmind.google/blog/page/8/	lab's own pages
Gemini 2.0 (Flash, Flash-Lite)Chips, models and products	The Gemini generation Google DeepMind introduced for the "agentic era", with Flash and Flash-Lite models for developers. released (superseded)Key facts Announced December 2024; generally available February 2025; "Start building with Gemini 2.0 Flash and Flash-Lite" February 2025; native image generation experiment March 2025. Model cards: Gemini 2.0 Flash (updated 15 Apr 2025), Flash-Lite (10 Apr 2025); Gemini 2.0 Pro released as experimental only (model-cards footnote).Made by Google DeepMind · 2024-12	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/7/ · deepmind.google/blog/page/7/	lab's own pages
Gemini 2.5 (Pro, Flash, Flash-Lite, Deep Think)Chips, models and products	Gemini's "thinking" model generation of 2025; an advanced Deep Think version reached gold-medal standard at the International Mathematical Olympiad. released (superseded by Gemini 3)Key facts Gemini 2.5 announced March 2025; 2.5 Flash April 2025; 2.5 Pro I/O updates May 2025; family expanded June 2025; 2.5 Flash-Lite generally available July 2025; Deep Think version "officially achieves gold-medal standard at the International Mathematical Olympiad" July 2025; Deep Think in the Gemini app August 2025; native audio June 2025. Model cards: 2.5 Pro (27 Jun 2025), 2.5 Flash and Flash Image (26 Sep 2025), 2.5 Flash-Lite (26 Sep 2025), 2.5 Deep Think (1 Aug 2025).Made by Google DeepMind · 2025-03Read the paper	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/7/ · deepmind.google/blog/page/6/	lab's own pages
Gemini 3 FlashChips, models and products	A fast, lower-cost Gemini 3 model. releasedKey facts Released 17 Dec 2025; model card same date. Successor Gemini 3.5 Flash "is based on Gemini 3 Flash" (19 May 2026).Made by Google DeepMind · 2025-12-17	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml · deepmind.google/models/model-cards/	lab's own pages
Gemini 3 ProChips, models and products	The first Gemini 3 model, launched in November 2025 for consumers and developers. released (superseded by 3.1 Pro)Key facts Announced November 2025 ("A new era of intelligence with Gemini 3"; "Start building with Gemini 3"); model card updated 18 Nov 2025. Successor Gemini 3.1 Pro "is based on" Gemini 3 Pro (19 Feb 2026).Made by Google DeepMind · 2025-11-18	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/5/ · deepmind.google/models/model-cards/	lab's own pages
Gemini 3.1 Flash-LiteChips, models and products	The fastest, cheapest Gemini 3 series model at its release. releasedKey facts Released 3 Mar 2026; model card same date. Followed by Gemini 3.5 Flash-Lite (21 Jul 2026).Made by Google DeepMind · 2026-03-03	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml · deepmind.google/models/model-cards/	lab's own pages
Gemini 3.1 ProChips, models and products	Google DeepMind's February 2026 Pro model for complex reasoning and coding tasks. released (API and apps)Key facts Model card 19 Feb 2026: based on Gemini 3 Pro; up to 1M tokens input, 64K output; Gemini app, Vertex AI, AI Studio, Gemini API, Antigravity, Gemini Enterprise, NotebookLM. Benchmarks (Thinking, High; as published Feb 2026): Humanity's Last Exam 44.4% (no tools), 51.4% (search and code); ARC-AGI-2 77.1%; GPQA Diamond 94.3%; SWE-Bench Verified 80.6%; SWE-Bench Pro 54.2%; Terminal-Bench 2.0 68.5%; BrowseComp 85.9%.Made by Google DeepMind · 2026-02-19	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/models/model-cards/gemini-3-1-pr · deepmind.google/blog/rss.xml	lab's own pages
Gemini 3.5 FlashChips, models and products	The first Gemini 3.5 model, built for complex, multi-step agent tasks. released (apps, API, Antigravity)Key facts Gemini 3.5 announced 15 May 2026; model card 19 May 2026: based on Gemini 3 Flash; 1M tokens input, 64K output. Benchmarks (May 2026, as published): Terminal-bench 2.1 76.2%; SWE-Bench Pro 55.1%; MCP Atlas 83.6%; Toolathlon 56.5%; OSWorld-Verified 78.4%; Finance Agent v2 57.9%; CharXiv Reasoning 84.2%; ARC-AGI-2 72.1%. Computer use added 24 Jun 2026.Made by Google DeepMind · 2026-05-15	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/models/model-cards/gemini-3-5-fl · deepmind.google/blog/rss.xml	lab's own pages
Gemini 3.5 Flash Cyber (and 3.8 Flash Cyber)Chips, models and products	A lightweight Gemini model for finding and patching software security flaws. releasedKey facts 3.5 Flash Cyber 17 Jul 2026; 3.8 Flash Cyber 2 Sep 2026 alongside 3.8 Flash; related post "Proactive cyber defense for governments and enterprises" (2 Sep 2026).Made by Google DeepMind · 2026-07-17	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml · deepmind.google/blog/rss.xml	lab's own pages
Gemini 3.6 Flash and Gemini 3.5 Flash-LiteChips, models and products	Mid-2026 Gemini Flash and Flash-Lite models for faster, cheaper tasks. releasedKey facts Released 21 Jul 2026 (with 3.5 Flash Cyber); model cards for both updated 21 Jul 2026.Made by Google DeepMind · 2026-07-21	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml · deepmind.google/models/model-cards/	lab's own pages
Gemini 3.7 FlashChips, models and products	An August 2026 Gemini Flash model. releasedKey facts Released 13 Aug 2026; model card same date.Made by Google DeepMind · 2026-08-13	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml · deepmind.google/models/model-cards/	lab's own pages
Gemini 3.8 FlashChips, models and products	Google DeepMind's current Flash model, aimed at complex agent tasks at scale. releasedKey facts Released 2 Sep 2026 with 3.8 Flash Cyber; model card same date; models page (read 30 Oct 2026): "Best for tackling complex agentic tasks at scale".Made by Google DeepMind · 2026-09-02	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/models/ · deepmind.google/models/model-cards/	lab's own pages
Gemini 4 ArgonChips, models and products	The first Gemini 4 model, aimed at software engineering, law, finance and defensive security work. released (apps, API, AI Studio, Antigravity)Key facts Announced 30 Sep 2026 ("our next era of frontier intelligence"); released 13 Oct 2026 with model card and safety post. Card: not a fine-tune of a prior model; text, image, audio and video input; up to 1M-token context; text output. Benchmarks (Oct 2026, as published): Vals Index 68.9%; AutomationBench 51.3%; DeepSWE v1.1 77.9%; Vibe Code Bench 91.9%; Terminal-bench 4.0 57.4%; PostTrainBench 45.3%; LABBench 2 88.8%; GraphWalks BFS 256k–1M 84.2%.Made by Google DeepMind · 2026-10-13	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/models/model-cards/gemini-4-argo · deepmind.google/blog/rss.xml	lab's own pages
Gemini 4 Argon for RoboticsChips, models and products	A robotics version of Gemini 4 Argon. releasedKey facts Announced 13 Oct 2026 (feed; developers blog link on the models page). Details page not opened. Earlier robotics line: Gemini Robotics 2 (28 Jul 2026), Robotics-ER 2 (30 Jul 2026).Made by Google DeepMind · 2026-10-13	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml	one source
Gemini 4 CarbonChips, models and products	Google DeepMind's newest frontier Gemini model, announced two days before this ledger's cut-off. released (announced)Key facts Announced 28 Oct 2026; listed at the top of the models page (read 30 Oct 2026). Launch post redirects to blog.google and could not be opened, so access and benchmarks are not recorded.Made by Google DeepMind · 2026-10-28	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml · deepmind.google/models/	lab's own pages
Gemini DiffusionChips, models and products	Experimental Google DeepMind language model that writes text by diffusion, refining whole passages at once rather than word by word. research (experimental demo)Key facts Sampling speed stated as 1,479 tokens per second excluding overhead (overhead 0.84 seconds), averaged across reported evaluations; access: experimental demo with sign-up for updates; DiffusionGemma (BIZ-LAB-7161) is stated to build on its researchMade by Google DeepMind · — (release date not stated on the model page)	London	deepmind.google/models/gemini-diffusion/	one source
Gemini EmbeddingChips, models and products	A Gemini model that turns text and other media into numbers for search and retrieval. released (date not on page)Key facts Theme: General-purpose AI models. Listed on the models page (read 30 Oct 2026); no date shown.Made by Google DeepMind	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/models/	one source
Gemini Robotics 1.5 and Gemini Robotics-ER 1.5Chips, models and products	A robot action model that reasons before acting, paired with a planning model that can call tools such as Google Search. released (ER 1.5 via Gemini API; 1.5 to select partners)Key facts Announced 25 Sep 2025. Access: Gemini Robotics-ER 1.5 through the Gemini API in Google AI Studio; Gemini Robotics 1.5 to select partners. Evaluated on 15 academic embodied-reasoning benchmarks.Made by Google DeepMind · 2025-09-25	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/gemini-robotics-15-brings-a · deepmind.google/models/model-cards/	lab's own pages
Gemini Robotics 2 (with Robotics-ER 2 and On-Device 2)Chips, models and products	Google DeepMind's second-generation robot models, adding whole-body control, video understanding and multi-robot teamwork. releasedKey facts Gemini Robotics 2 announced 28 Jul 2026; Gemini Robotics ER 2 30 Jul 2026 (cards for ER 2 and On-Device 2 dated 30 Jul 2026); models page: "The intelligence layer to power any kind of robot". Related: "Powering the future of robotics in Europe" (9 Jun 2026).Made by Google DeepMind · 2026-07-28	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml · deepmind.google/blog/rss.xml	lab's own pages
Gemini Robotics and Gemini Robotics-ERChips, models and products	Gemini-based models that let robots act on what they see and are told; the ER version adds spatial reasoning for robot builders. released (trusted testers; Apptronik partnership)Key facts Announced 12 Mar 2025; built on Gemini 2.0. Figure: "more than doubles performance on a comprehensive generalization benchmark". Access: Gemini Robotics-ER to trusted testers; partnership with Apptronik on humanoid robots. Version history: Gemini Robotics (Mar 2025) → On-Device (Jun 2025) → 1.5 (Sep 2025) → ER 1.6 (Apr 2026) → Gemini Robotics 2 / ER 2 / On-Device 2 (Jul 2026) → Gemini 4 Argon for Robotics (Oct 2026).Made by Google DeepMind · 2025-03-12Versions Gemini Robotics (Mar 2025) → On-Device (Jun 2025) → 1.5 (Sep 2025) → ER 1.6 (Apr 2026) → Gemini Robotics 2 / ER 2 / On-Device 2 (Jul 2026) → Gemini 4 Argon for Robotics (Oct 2026)	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/gemini-robotics-brings-ai-i · deepmind.google/blog/page/7/	one source
Gemini Robotics On-DeviceChips, models and products	A robotics model for two-armed robots that runs on the robot itself, without a network connection, and can be fine-tuned. released (SDK for trusted testers)Key facts Announced 24 Jun 2025; Gemini Robotics SDK for trusted testers. Figure: adapts to new tasks with "50 to 100 demonstrations". Model card updated 1 Jul 2025; Gemini Robotics On-Device 2 card dated 30 Jul 2026.Made by Google DeepMind · 2025-06-24	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/gemini-robotics-on-device-b · deepmind.google/models/model-cards/	lab's own pages
Gemini Robotics-ER 1.6Chips, models and products	A robotics reasoning model with better spatial and multi-camera understanding. releasedKey facts Released 13 Apr 2026; model card updated 20 Apr 2026.Made by Google DeepMind · 2026-04-13	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml · deepmind.google/models/model-cards/	lab's own pages
Gemma (first generation)Chips, models and products	Google's first family of lightweight open-weight models, built from the technology behind Gemini. released (open weights)Key facts Theme: General-purpose AI models. Announced February 2024 (blog index). Gemma page (read 30 Oct 2026): "a collection of lightweight, open models built from the same technology that powers our Gemini models"; downloadable via Kaggle, Hugging Face, Ollama and others. Variants with model cards: Gemma 1, Gemma 2, CodeGemma, RecurrentGemma, PaliGemma 1 and 2, ShieldGemma 1 and 2 (cards updated 25 Feb 2025).Made by Google DeepMind · 2024-02Versions Gemma → Gemma 2 (and CodeGemma, RecurrentGemma, PaliGemma → PaliGemma 2, ShieldGemma → ShieldGemma 2, DataGemma)	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/9/ · deepmind.google/models/gemma/	lab's own pages
Gemma 2 (2B, 9B, 27B; plus Gemma 2 JPN)Chips, models and products	Second generation of Google DeepMind's small open-weight language models, released in three sizes. released (open weights)Key facts 9B and 27B released 27 Jun 2024; 2B released 31 Jul 2024; Gemma 2 JPN (2B) released 3 Oct 2024; 27B trained on 13 trillion tokens, 9B on 8 trillion, 2B on 2 trillion (model card); trained on TPUv5p; access: open weightsMade by Google DeepMind · 2024-06-27Versions Gemma → Gemma 2 (and CodeGemma, RecurrentGemma, PaliGemma → PaliGemma 2, ShieldGemma → ShieldGemma 2, DataGemma)	London	ai.google.dev/gemma/docs/releases · ai.google.dev/gemma/docs/core/model_card_2	lab's own pages
Gemma 3 (incl. Gemma 3 270M)Chips, models and products	The third generation of Google's open-weight Gemma models; a tiny 270-million-parameter version followed in August 2025. released (open weights)Key facts Theme: General-purpose AI models. "Introducing Gemma 3" March 2025; "Introducing Gemma 3 270M: The compact model for hyper-efficient AI" August 2025; model card updated 14 Aug 2025; Gemma Scope 2 interpretability tools cover the Gemma 3 family (16 Dec 2025); a Gemma model helped identify a potential cancer therapy pathway (October 2025 index).Made by Google DeepMind · 2025-03	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/7/ · deepmind.google/blog/page/6/	lab's own pages
Gemma 3nChips, models and products	An open Gemma model variant introduced in June 2025 with a developer guide. released (open weights)Key facts Theme: General-purpose AI models. "Introducing Gemma 3n: The developer guide" June 2025; model card updated 17 Jun 2025.Made by Google DeepMind · 2025-06	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/6/ · deepmind.google/models/model-cards/	lab's own pages
Gemma 4 (E2B, E4B, 12B, 26B, 31B)Chips, models and products	Google DeepMind's fourth generation of open-weight models, built for reasoning and agent workflows. released (open weights)Key facts Theme: General-purpose AI models. Released 2 Apr 2026; sizes E2B, E4B, 12B, 26B, 31B (Gemma page); Gemma 4 QAT June 2026; Gemma 4 12B "a unified, encoder-free multimodal model" 9 Jun 2026; model card updated 2 Apr 2026. Downloads via Kaggle, Hugging Face, Ollama, LM Studio and others. Launch post redirects to blog.google and was not opened (no download figures recorded).Made by Google DeepMind · 2026-04-02	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml · deepmind.google/models/gemma/	lab's own pages
Genie (first version, "Generative Interactive Environments")Chips, models and products	First Genie world model: generates interactive, controllable environments, learned without labels from internet videos. research (paper only)Key facts arXiv 2402.15391 submitted 23 Feb 2024; 11B parameters; trained unsupervised on unlabelled internet videos; access: paper only (successors Genie 2, 3 and 4 already listed)Made by Google DeepMind · 2024-02-23Versions Genie → Genie 2 → 3 → 4	London	arxiv.org/abs/2402.15391	one source
Genie 2Chips, models and products	A world model that turns one image into a playable 3D world a person or AI agent can move through. research onlyKey facts Announced 4 Dec 2024. Figure: consistent worlds for up to a minute. Version history: Genie 1 (2D worlds) → Genie 2 (Dec 2024) → Genie 3 (Aug 2025) → Genie 4 (Oct 2026).Made by Google DeepMind · 2024-12-04Versions Genie 1 (2D worlds) → Genie 2 (Dec 2024) → Genie 3 (Aug 2025) → Genie 4 (Oct 2026)	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/genie-2-a-large-scale-found · deepmind.google/blog/page/8/	one source
Genie 3Chips, models and products	A world model that builds explorable, interactive virtual worlds from a text description in real time. limited research previewKey facts Announced 5 Aug 2025; limited preview for a small set of academics and creators. Figures (5 Aug 2025): 720p at 24 frames per second; consistent for a few minutes. Predecessors: Genie 1, Genie 2.Made by Google DeepMind · 2025-08-05	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/genie-3-a-new-frontier-for- · deepmind.google/blog/page/6/	one source
Genie 4Chips, models and products	Google DeepMind's newest world model, generating interactive worlds in real time with sound, physics and characters steered by text. research preview (trusted testers; waitlist)Key facts Announced 27 Oct 2026; research preview for trusted testers; coming to Google AI Ultra subscribers in Project Genie and to early-access Gemini API and AI Studio partners. Limitation stated: "a few minutes of continuous interaction". First Genie to generate audio natively.Made by Google DeepMind · 2026-10-27	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/genie-4-a-new-generation-of · deepmind.google/models/genie/	lab's own pages
Google DeepMindCompanies	Builds general-purpose AI models and research systems for Google and for science, including the Gemini models and the Genie world-simulation models. merged Profile Founded 2010 (DeepMind); Google DeepMind formed 20 April 2023 when DeepMind and the Brain team from Google Research merged Founders Demis Hassabis — co-founder and CEO (other co-founders not named on pages opened) Chief executive Demis Hassabis — CEO of the combined Google DeepMind from 20 April 2023 Chief scientist or research lead Koray Kavukcuoglu — led the new Scientific Board, with the research leadership team reporting to him (April 2023; current title not checked) Other leaders Eli Collins — VP of Product (April 2023); Zoubin Ghahramani — research leadership team, reporting to Koray Kavukcuoglu (April 2023); Lila Ibrahim, Pushmeet Kohli, Owen Larter — named co-authors with Demis Hassabis of the 10 December 2025 UK partnership post (roles not stated on that page) Hq null (careers page lists London first but does not call it headquarters); other_offices: Bay Area, Bangalore, Cambridge (US), Montreal, New York City, Paris, Tokyo, Toronto, Zurich (careers page, read 30 Oct 2026) What it does Alphabet's AI research unit, formed in 2023 by merging DeepMind with Google's Brain team. It builds the Gemini family of models, generative media models (Veo for video, Lyria for music, Nano Banana for images, Genie for interactive worlds), Gemini Robotics models, the open Gemma models, and science systems such as AlphaFold, AlphaGenome, AlphaEvolve and WeatherNext. Research focus "build AI responsibly to benefit humanity"; areas listed on its site: frontier models (Gemini), generative media, world models, robotics, games and agents (SIMA), AI for science (AlphaFold, AlphaGenome, WeatherNext, AlphaEarth, AlphaEvolve, Co-Scientist), frontier safety Models and products Gemini 4 Carbon — "Our next frontier model" — October 2026; Gemini 4 Argon (and Gemini 4 Argon for Robotics) — October 2026; Genie 4 — world simulation model — October 2026; EmbeddingGemma 2 — open multimodal embedding model — October 2026; Gemini Robotics 2 — robot "intelligence layer" — 30 July 2026; Gemini Robotics On-Device 2 — runs locally on robots — 30 July 2026; Gemini Robotics ER 2 — embodied reasoning model — July 2026; Gemini Robotics-ER 1.6 — April 2026; Gemini 3.8 Flash — "most intelligent workhorse model yet for coding and agents" — date not shown; Gemma 4 — open models — date not shown; Gemini Omni 1.1 Flash — date not shown; Lyria 3.5 — music model — date not shown; Nano Banana 2 Lite — image model — date not shown; Gemini Robotics 1.5 — September 2025; Gemini Robotics — March 2025 Customers or partners named UK government (i.AI), Google Cloud, Faculty, and Barnet, Dorset and Camden planning authorities — AI planning-officer prototype — 16 June 2026 — https://deepmind.google/blog/unlocking-uk-house-building-with-ai-accelerated-planning/; Apptronik, Boston Dynamics, Agile Robots — thanked for support on Gemini Robotics 2 — 30 July 2026 — https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/; Singapore Government — fine-tuned model for scam detection — October 2026 — https://deepmind.google/ Papers and awards AlphaFold described as "Nobel-winning" (Isomorphic Labs homepage; year not given on page) — https://isomorphiclabs.com/?amp=&amp= Funding total as published null (unit of Alphabet) Rounds found none (merger already held as DEAL-LAB-001) Uk government links Partnership announced 10 December 2025 — UK scientists get priority access to AlphaEvolve, AlphaGenome, AI co-scientist and WeatherNext; Google DeepMind's first automated laboratory, focused on materials science, to open in the UK in 2026; deeper work with the AI Security Institute; AI cyber-defence work (Big Sleep, CodeMender); education research on Gemini and England's national curriculum. Planning: i.AI's Gemini-based "Extract" made available to every council in England in June 2026; a national planning tool planned for all councils from 2027. Already on site as BIZ-BIG-2009, BIZ-SAFE-2085 and BIZ-SVC-030 / DEAL-SVC-013. Status now Operating as Alphabet's AI unit; its homepage lists Gemini 4 Carbon, Gemini 4 Argon and Genie 4 among October 2026 news. Labs Round 7 deep profile (30 Oct 2026), as the researchers wrote it; confidence: single_source (company pages only). Sources deepmind.google/about/ deepmind.google/blog/announcing-google-deepmind/ deepmind.google/blog/strengthening-our-partnersh deepmind.google/blog/unlocking-uk-house-building deepmind.google/blog/gemini-robotics-2-brings-wh deepmind.google/ deepmind.google/careers/ Models and products (140) Version chainsLyria → Lyria 2Gemini Robotics (Mar 2025) → On-Device (Jun 2025) → 1.5 (Sep 2025) → ER 1.6 (Apr 2026) → Gemini Robotics 2 / ER 2 / On-Device 2 (Jul 2026) → Gemini 4 Argon for Robotics (Oct 2026)Gemma → Gemma 2 (and CodeGemma, RecurrentGemma, PaliGemma → PaliGemma 2, ShieldGemma → ShieldGemma 2, DataGemma)Genie 1 (2D worlds) → Genie 2 (Dec 2024) → Genie 3 (Aug 2025) → Genie 4 (Oct 2026)Veo 2 (Dec 2024) → Veo 3 → Veo 3.1 (Oct 2025) → Veo 3.1 Lite (card 8 Apr 2026)AlphaGeometry → AlphaGeometry 2Genie → Genie 2 → 3 → 4AlphaCode → AlphaCode 2AlphaGo → AlphaZero → MuZeroAlphaGo Fan → AlphaGo Lee → AlphaGo Master → AlphaGo Zero Gemini 4 Carbon 2026-10-28 released (announced)Google DeepMind's newest frontier Gemini model, announced two days before this ledger's cut-off. Genie 4 2026-10-27 research preview (trusted testers; waitlist)Google DeepMind's newest world model, generating interactive worlds in real time with sound, physics and characters steered by text. Gemini 4 Argon 2026-10-13 released (apps, API, AI Studio, Antigravity)The first Gemini 4 model, aimed at software engineering, law, finance and defensive security work. Gemini 4 Argon for Robotics 2026-10-13 releasedA robotics version of Gemini 4 Argon. SynthID Bio 2026-09-30 proof of conceptA proof of concept for watermarking AI-designed proteins without changing how they work. Gemini 3.8 audio models (Live, Live Extended Thinking, TTS, Live Avatar) 2026-09-15 releasedGemini's September 2026 real-time voice, text-to-speech and talking-avatar models. AlphaGenome Atlas 2026-09-08 released (free portal for academic research; API)A free catalogue predicting the molecular effects of about 9 billion possible single-letter changes in the human genome. Gemini 3.8 Flash 2026-09-02 releasedGoogle DeepMind's current Flash model, aimed at complex agent tasks at scale. Gemini 3.7 Flash 2026-08-13 releasedAn August 2026 Gemini Flash model. Lyria 3.5 2026-07-29 released (Google Flow Music; Gemini)Google DeepMind's newest music model, with better musicality, lyrics and vocals. Gemini Robotics 2 (with Robotics-ER 2 and On-Device 2) 2026-07-28 releasedGoogle DeepMind's second-generation robot models, adding whole-body control, video understanding and multi-robot teamwork. Gemini 3.6 Flash and Gemini 3.5 Flash-Lite 2026-07-21 releasedMid-2026 Gemini Flash and Flash-Lite models for faster, cheaper tasks. Gemini 3.5 Flash Cyber (and 3.8 Flash Cyber) 2026-07-17 releasedA lightweight Gemini model for finding and patching software security flaws. Nano Banana 2 Lite (Gemini 3.1 Flash-Lite Image) and Nano Banana 2.1 2026-06-30 releasedLighter and updated versions of the Nano Banana 2 image model. DiffusionGemma 2026-06-10 released (open weights)An open Gemma model that generates text by diffusion, which the lab says is four times faster. Gemini 3.5 Live Translate and Gemini 3.5 Transcribe 2026-06-09 releasedGemini speech models for near real-time spoken translation and for speech-to-text transcription. AlphaFold Protein Structure Database 2026-05-19 activeA free public database of AI-predicted protein structures, made by Google DeepMind's AlphaFold and run with EMBL-EBI. Gemini Omni (Omni, Omni Flash, Omni 1.1 Flash) 2026-05-17 releasedGoogle DeepMind's model for creating content "from anything", starting with video. Gemini 3.5 Flash 2026-05-15 released (apps, API, Antigravity)The first Gemini 3.5 model, built for complex, multi-step agent tasks. Co-Scientist 2026-05-12 releasedA Gemini-based multi-agent AI partner built to help researchers speed up scientific work. AI co-clinician 2026-04-30 researchGoogle DeepMind research towards an AI that works alongside clinicians in patient care. Decoupled DiLoCo 2026-04-22 researchA Google DeepMind method for resilient training of AI across distributed computers. Gemini Robotics-ER 1.6 2026-04-13 releasedA robotics reasoning model with better spatial and multi-camera understanding. Gemma 4 (E2B, E4B, 12B, 26B, 31B) 2026-04-02 released (open weights)Google DeepMind's fourth generation of open-weight models, built for reasoning and agent workflows. Gemini 3.1 Flash Live and Gemini 3.1 Flash TTS 2026-03-26 releasedGemini voice models: Flash Live for real-time spoken conversation and Flash TTS for directed, expressive speech. Lyria 3 Pro 2026-03-25 releasedA Lyria music model that writes longer tracks with song structure. Gemini 3.1 Flash-Lite 2026-03-03 releasedThe fastest, cheapest Gemini 3 series model at its release. Nano Banana 2 (Gemini 3.1 Flash Image) 2026-02-26 releasedA fast Gemini image generation model with Pro-level features. Gemini 3.1 Pro 2026-02-19 released (API and apps)Google DeepMind's February 2026 Pro model for complex reasoning and coding tasks. Lyria 3 2026-02-18 released (Gemini app, Vertex AI, AI Studio, YouTube, API)Google DeepMind's music generation model that turns a text prompt into a song; it made music creation available in the Gemini app. Gemini 3 Deep Think 2026-02-12 released (updated reasoning mode)Gemini's most specialised reasoning mode, updated for science, research and engineering problems. Project Genie 2026-01-29 released (prototype for Google AI Ultra subscribers in the US)An experimental app in which users create and explore AI-generated interactive worlds. D4RT 2026-01-16 researchA model for reconstructing and tracking scenes in four dimensions (3D plus time), claimed to be much faster than earlier methods. Gemini 3 Flash 2025-12-17 releasedA fast, lower-cost Gemini 3 model. Gemma Scope 2 2025-12-16 released (open tools)Open interpretability tools that let researchers inspect the inner workings of every Gemma 3 model. FunctionGemma and TranslateGemma 2025-12 released (open weights)Specialised open Gemma models: one for calling software functions, one for translation. Nano Banana Pro (Gemini 3 Pro Image) 2025-11-20 released (Gemini app and API)Google DeepMind's Gemini 3 Pro image generation and editing model. Gemini 3 Pro 2025-11-18 released (superseded by 3.1 Pro)The first Gemini 3 model, launched in November 2025 for consumers and developers. SIMA 2 2025-11-13 limited research previewA Gemini-powered game agent that follows instructions, reasons about goals, talks with users and improves itself in 3D worlds. Google Antigravity (and Antigravity 2.0) 2025-11 released (product)A Google product, launched November 2025, through which Gemini models and DeepMind tools (such as an AlphaGenome skill) are offered; version 2.0 followed in May 2026. WeatherNext 2 2025-11 releasedGoogle DeepMind's November 2025 weather forecasting model. CodeMender 2025-10-06 research (not public; human-reviewed patches)An AI agent that finds and fixes security flaws in software and rewrites code to make it safer. Gemini 2.5 Computer Use 2025-10 released (model card)A Gemini model that operates computer interfaces such as web pages by clicking and typing. Veo 3 and Veo 3.1 (incl. Veo 3.1 Lite) 2025-10 released (Gemini app, Flow, Gemini API, AI Studio, Vids)Google DeepMind's leading video model, which makes 8-second clips at 1080p or 4K. Gemini Robotics 1.5 and Gemini Robotics-ER 1.5 2025-09-25 released (ER 1.5 via Gemini API; 1.5 to select partners)A robot action model that reasons before acting, paired with a planning model that can call tools such as Google Search. VaultGemma 2025-09 released (open weights)A Gemma language model trained with differential privacy, a privacy-protecting training method. EmbeddingGemma (and EmbeddingGemma 2) 2025-09 released (open weights)Small open Gemma models that turn text (and, in version 2, other media) into numbers for search and retrieval. Genie 3 2025-08-05 limited research previewA world model that builds explorable, interactive virtual worlds from a text description in real time. Gemini 2.5 Flash Image 2025-08 releasedA Gemini image generation and editing model; the models page brands Gemini Image as "Nano Banana". AlphaEarth Foundations 2025-07-30 released (Satellite Embedding dataset in Google Earth Engine)An AI model that combines satellite and other Earth data into one map-like representation for tracking land and environment. Aeneas 2025-07-23 released (free tool; open code and data)An AI model that helps historians date, place and restore ancient Latin inscriptions. MedGemma (incl. MedGemma 1.5 4B) 2025-07 released (open weights)Open Gemma-based models for building health AI applications. T5Gemma (and T5Gemma 2) 2025-07 released (open weights)A collection of Gemma models with an encoder-decoder design. AlphaGenome 2025-06-25 released (API preview for non-commercial research)An AI model that predicts how single DNA changes affect the processes that switch genes on and off. Gemini Robotics On-Device 2025-06-24 released (SDK for trusted testers)A robotics model for two-armed robots that runs on the robot itself, without a network connection, and can be fine-tuned. Gemma 3n 2025-06 released (open weights)An open Gemma model variant introduced in June 2025 with a developer guide. Imagen 4 2025-05-20 releasedGoogle DeepMind's Imagen 4 image generation model. AlphaEvolve 2025-05-14 released (early access for academics)Google DeepMind's coding agent that uses Gemini models plus automatic checkers to discover and improve algorithms. SynthID Detector 2025-05 released (portal)A portal to help identify AI-generated content marked with SynthID. Lyria 2 (and Music AI Sandbox update) 2025-04-24 released (waitlist, US)Second Lyria music model, offered to musicians and producers in the US through Google's Music AI Sandbox. DolphinGemma 2025-04 research releaseA Gemma-based model used to help decode dolphin communication. Gemini Robotics and Gemini Robotics-ER 2025-03-12 released (trusted testers; Apptronik partnership)Gemini-based models that let robots act on what they see and are told; the ER version adds spatial reasoning for robot builders. ShieldGemma 2 2025-03-10 released (open weights)Open model that checks images for harmful content against set safety policies. Gemini 2.5 (Pro, Flash, Flash-Lite, Deep Think) 2025-03 released (superseded by Gemini 3)Gemini's "thinking" model generation of 2025; an advanced Deep Think version reached gold-medal standard at the International Mathematical Olympiad. Gemma 3 (incl. Gemma 3 270M) 2025-03 released (open weights)The third generation of Google's open-weight Gemma models; a tiny 270-million-parameter version followed in August 2025. Gemini 2.0 Flash Thinking (experimental) 2024-12-19 superseded (public preview)First Gemini model released in a "thinking" mode that reasons step by step before answering. PaliGemma 2 (and PaliGemma 2 mix) 2024-12-05 released (open weights)Second version of the open image-and-text Gemma model, in three sizes. Genie 2 2024-12-04 research onlyA world model that turns one image into a playable 3D world a person or AI agent can move through. GenCast 2024-12-04 released (open code and weights)An AI weather model that gives a range of possible forecasts up to 15 days ahead, including extreme-weather risk. Gemini 2.0 (Flash, Flash-Lite) 2024-12 released (superseded)The Gemini generation Google DeepMind introduced for the "agentic era", with Flash and Flash-Lite models for developers. Veo 2 2024-12 released (superseded)Google DeepMind's second video generation model, released with Imagen 3 in December 2024. Imagen 3 2024-12 released (superseded)Google DeepMind's Imagen 3 image generation model, announced with Veo 2. FACTS Grounding and FACTS Benchmark Suite 2024-12 released (benchmarks)Google DeepMind benchmarks that test how factually accurate language models are. AlphaQubit 2024-11 research releaseAn AI system aimed at quantum error correction (per its post address). Multi-speaker dialogue generation (behind NotebookLM Audio Overviews) 2024-10-30 live (inside Google products)Google DeepMind speech technology that writes out two-person podcast-style conversations as audio, used in NotebookLM Audio Overviews. AlphaChip 2024-09-26 released (open checkpoint and weights)A reinforcement-learning method that lays out computer chips, used for Google's TPU designs. Veo (first version) in YouTube Dream Screen 2024-09-18 (YouTube integration announced; May 2024 launch post on blog.google not opened) superseded (product integration)Google DeepMind's first Veo video model, brought to YouTube Shorts to make short background clips from a chosen image. DataGemma (RIG and RAG) 2024-09-12 released (open weights; research tool)Gemma 2 models tuned to answer plain-language questions with public statistics drawn from Google's Data Commons. ALOHA Unleashed and DemoStart 2024-09-12 research (paper only)Two robot-learning methods: one teaches two-armed robots complex tasks, the other trains dexterous hands in simulation. AlphaProteo 2024-09-05 released (paper only)An AI system that designs new proteins that bind tightly to chosen target molecules. Gemma Scope 2024-07-31 released (open sparse autoencoders)Open tools that let safety researchers look inside Gemma 2 models to see how they process information. ShieldGemma 2024-07-31 released (open weights)Open Gemma 2 models that check text going into and out of AI systems for harmful content. AlphaProof 2024-07-25 research (paper; competition result)Reinforcement-learning system that proves maths statements in the formal language Lean; solved three IMO 2024 problems. AlphaGeometry 2 2024-07-25 research (competition result)Improved version of AlphaGeometry, pairing a Gemini-based language model with a symbolic engine to solve olympiad geometry. Gemma 2 (2B, 9B, 27B; plus Gemma 2 JPN) 2024-06-27 released (open weights)Second generation of Google DeepMind's small open-weight language models, released in three sizes. V2A (video-to-audio research) 2024-06-17 research (not released)Research system that generates soundtracks, effects and dialogue to match the action in a video. PaliGemma 2024-05-14 released (open weights)Small open model that reads images and text together, for tasks such as captioning and answering questions about pictures. AlphaFold 3 and AlphaFold Server 2024-05-08 released (free server; code and weights for academic use)The third AlphaFold, which predicts the structures and interactions of proteins and other molecules, offered through a free web server. Med-Gemini 2024-04-29 research (paper only)Family of Gemini models specialised for medicine, tested on medical exam and imaging benchmarks. CodeGemma 2024-04-09 released (open weights)Small open Gemma models trained to complete, write and chat about computer code. RecurrentGemma (2B and 9B) 2024-04-09 released (open weights)Open Gemma language model built on a different design (Griffin) that mixes recurrence with local attention to run more efficiently. TacticAI 2024-03-19 released (paper; Liverpool FC collaboration)An AI assistant that advises football coaches on corner kicks by predicting outcomes and suggesting player positions. SIMA (Scalable Instructable Multiworld Agent) 2024-03-13 research only (technical report)An AI agent that follows spoken-style instructions to carry out tasks in 3D video games using keyboard and mouse. Genie (first version, "Generative Interactive Environments") 2024-02-23 research (paper only)First Genie world model: generates interactive, controllable environments, learned without labels from internet videos. Gemini 1.5 2024-02 released (superseded)The second Gemini generation; a May 2024 update added a faster model and a longer context window. Gemma (first generation) 2024-02 released (open weights)Google's first family of lightweight open-weight models, built from the technology behind Gemini. AlphaGeometry 2024-01-17 released (paper; open code and model)A system that pairs a language model with a logic engine to prove Olympiad geometry problems. FunSearch 2023-12-14 released (paper)A system in which a language model writes programs and an automatic checker scores them, used to find new maths results. Gemini 1.0 2023-12 released (superseded)Google DeepMind's first Gemini model family, described at launch as its largest and most capable AI model. GNoME (Graph Networks for Materials Exploration) 2023-11-29 released (paper and data)A deep-learning tool that predicts which new crystal materials are stable. Lyria (first version) and Dream Track 2023-11-16 released (limited: YouTube experiment)Google DeepMind's first Lyria music-generation model, launched with a YouTube Shorts experiment that made songs in artists' voices. GraphCast 2023-11-14 released (open code)An AI weather model that makes 10-day global forecasts in under a minute. Open X-Embodiment dataset and RT-X models (RT-1-X, RT-2-X) 2023-10-03 released (open dataset; models)Pooled robot-training dataset from many labs and the robot models trained on it, which work across different robot types. AlphaMissense 2023-09-19 released (open catalogue CC BY 4.0; open code)An AI model that estimates whether single-letter changes in human proteins are likely to cause disease. SynthID 2023-08-29 released (in Google products; text watermarking open-source planned May 2024)An invisible watermark that marks AI-generated images, then text, video and audio, so they can be identified later. RT-2 (Robotic Transformer 2) 2023-07-28 released (paper)A robot-control model that learns from web and robot data and turns camera images and instructions into robot actions. RoboCat 2023-06-20 research (paper only)Self-improving robot agent that learns new tasks on different real robot arms from a few hundred demonstrations. AlphaDev 2023-06-07 released (paper; algorithms open-sourced into LLVM libc++)An AI system that found faster sorting and hashing routines by building programs one low-level instruction at a time. Imagen 2 2023 (exact day not stated on source) superseded (released)Second version of Google's Imagen text-to-image model, with an added model for judging image aesthetics. AlphaCode 2 2023 (reported in DeepMind review dated 22 Dec 2023) research (technical report)Second AlphaCode system for competitive programming, solving more contest problems than the first. DeepNash (Stratego agent) 2022-12-01 research (Science paper)DeepMind agent that taught itself the hidden-information board game Stratego to human expert level. AlphaTensor 2022-10-05 released (paper and code)An AI system that searches for faster ways to do matrix multiplication and checks they are correct. Sparrow (dialogue agent) 2022-09-22 research (paper only)Research chatbot trained with human feedback and rules to give evidence-backed answers and avoid unsafe replies. Gato 2022-05-12 released (paper)One neural network with one set of weights that plays games, captions images, chats and controls a robot arm. MuJoCo (open-sourced physics simulator) 2022-05 released (open source)A physics simulator for robotics research that DeepMind opened up in 2021 and open-sourced in 2022. Flamingo 2022-04-28 released (paper)A single model that reads images, video and text together and learns new tasks from a few examples. Chinchilla 2022-04-12 released (paper)A 70-billion-parameter language model that showed training on more data can beat simply adding parameters. Ithaca 2022-03 released (paper)An AI model for ancient Greek inscriptions; the predecessor of Aeneas. AlphaCode 2022-02-02 released (paper and dataset)A system that writes computer programs for programming contests by generating many candidates and filtering them. Gopher 2021-12-08 released (paper)A 280-billion-parameter language model that DeepMind used to study scale, risks and retrieval. RETRO (Retrieval-Enhanced Transformer) 2021-12-08 research (paper only)Language model that looks up passages in a very large text database while it writes, so a smaller model can match much larger ones. Enformer 2021-10-04 released (open model; paper)Model that predicts how DNA sequence controls gene activity, reading long stretches of DNA at once. DGMR precipitation nowcasting model (with the Met Office) 2021-09-29 research (paper; Nature)Rain-forecasting model built with the Met Office that predicts the next 90 minutes of rainfall from radar. Perceiver and Perceiver IO 2021-08-03 released (open code; papers)General neural-network designs that take in many kinds of data (images, sound, text) and produce many kinds of output. XLand open-ended learning agents 2021-07-27 research (paper only)DeepMind agents trained across hundreds of thousands of generated 3D games that learned general game-playing behaviour. MuZero 2020-12-23 released (paper)A game-playing system that learns its own model of a game and plans with it, without being told the rules. AlphaFold2 2020-11-30 released (open code)The second AlphaFold, which predicts protein shapes close to lab-measured accuracy; organisers called it a solution to protein folding. FermiNet (Fermionic Neural Network) 2020-10-19 (first post; updated 22 Aug 2024 for excited-states Science paper) released (open code; papers)Neural network that calculates the quantum behaviour of electrons in molecules from first principles. Agent57 2020-03 released (paper)A DeepMind game-playing agent reported to outperform the human benchmark on Atari games. AlphaStar 2019-01-24 released (research result)An AI program that plays the full real-time strategy game StarCraft II and beat top professional players. AlphaZero 2018-12-06 released (paper)One learning system that teaches itself chess, shogi and Go from the rules alone, by playing against itself. AlphaFold (first version, CASP13) 2018-12-02 released (paper)The first AlphaFold, which predicts a protein's 3D shape from its sequence; it ranked first at the CASP13 benchmark in 2018. Streams (DeepMind Health app) 2017-11 released (NHS hospital app)A mobile app for NHS hospital clinicians; DeepMind published a post explaining why it did not use AI. AlphaGo Zero 2017-10-18 released (paper)A Go-playing system that taught itself the game from random moves by playing against itself, with no human game records. WaveNet (in Google Assistant) 2017-10 released (in product)DeepMind's raw-audio speech generation model (first published 2016), launched in Google Assistant in October 2017. WeatherNext 3 — (no date on page) releasedGoogle DeepMind's AI global weather model, which produces a new forecast every hour and feeds Google Search, Maps and Gemini. Magenta RealTime and Lyria RealTime no date released (Magenta RealTime open; dates not on page)Real-time interactive music generation models; Magenta RealTime is the open Lyria model. Gemini Embedding no date released (date not on page)A Gemini model that turns text and other media into numbers for search and retrieval. Gemini Diffusion — (release date not stated on the model page) research (experimental demo)Experimental Google DeepMind language model that writes text by diffusion, refining whole passages at once rather than word by word. Project Astra — (date not stated on the project page) research prototype (trusted testers)Google DeepMind's prototype universal assistant that sees and hears through a phone camera or prototype glasses and talks back in real time. Selected papers (25) Nature · January 2026one sourceAlphaGenome predicts the molecular effect of changes to DNA and scores the effects of genetic variants by comparing mutated and unmutated sequences. Read it Nature · 12 Nov 2025one sourceDescribes AlphaProof, which together with a second DeepMind system solved four of the six problems at the 2024 International Mathematical Olympiad (silver-medal standard). Read it Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities technical report · 7 Jul 2025The report describes Gemini 2.5 Pro, then Google's strongest model, leading on frontier coding and reasoning tests. AlphaEvolve: A coding agent for scientific and algorithmic discovery arXiv preprint · 16 Jun 2025A coding agent found, among other results, the first improvement in 56 years on Strassen's matrix-multiplication method for one case. De novo design of high-affinity protein binders with AlphaProteo arXiv preprint · 12 Sep 2024AlphaProteo designed new proteins that bind target proteins 3 to 300 times more tightly than the best earlier methods. Accurate structure prediction of biomolecular interactions with AlphaFold 3 Nature · 8 May 2024joint paperOne model predicted how proteins and other biological molecules fit together, with high accuracy. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context technical report · 8 Mar 2024Gemini 1.5 could find facts almost perfectly across inputs of millions of tokens of text, video and audio. Genie: Generative Interactive Environments arXiv preprint · 23 Feb 2024Genie generates interactive virtual worlds that a user can act in frame by frame. Solving olympiad geometry without human demonstrations (AlphaGeometry) Nature · 17 Jan 2024AlphaGeometry solved 25 olympiad geometry test problems, against ten for the previous best method. GenCast: Diffusion-based ensemble forecasting for medium-range weather arXiv preprint · 25 Dec 2023A machine-learning ensemble forecaster was more skilful than the main operational ensemble (ENS) on 97.4% of 1,320 targets. Gemini: A Family of Highly Capable Multimodal Models technical report · 19 Dec 2023The first Gemini models, built to handle several kinds of input, set new best scores on 30 of 32 benchmarks reported. Mathematical discoveries from program search with large language models (FunSearch) Nature · 14 Dec 2023Pairing a language model with an automatic checker produced new mathematical constructions (larger cap sets) than were known before. Scaling deep learning for materials discovery (GNoME) Nature · 29 Nov 2023Graph neural networks trained at scale generalised well enough to speed up the discovery of new materials. Faster sorting algorithms discovered using deep reinforcement learning (AlphaDev) Nature · 7 Jun 2023AlphaDev found short sorting routines from scratch that beat the best known human-written ones. GraphCast: Learning skillful medium-range global weather forecasting arXiv preprint · 24 Dec 2022A machine-learning weather model beat the main operational forecast system on 90% of 1,380 test measures. Nature · 5 Oct 2022one sourceAlphaTensor searched for new ways to multiply matrices with fewer steps, finding, for one case, a method that needs just 76 multiplications. Read it A Generalist Agent (Gato) arXiv preprint · 12 May 2022A single network with one set of weights could play Atari games, caption images and chat, among other tasks. Training Compute-Optimal Large Language Models (Chinchilla) arXiv preprint · 29 Mar 2022For a fixed computing budget, model size and training data should grow together; most large language models of the time were undertrained. Competition-Level Code Generation with AlphaCode arXiv preprint (published in Science) · 8 Feb 2022AlphaCode wrote programs that ranked, on average, in the top 54.3% of human entrants in coding contests. Advancing mathematics by guiding human intuition with AI Nature · 1 Dec 2021Machine learning helped mathematicians spot patterns that led to new results in pure mathematics. Highly accurate protein structure prediction with AlphaFold (AlphaFold 2) Nature · 15 Jul 2021AlphaFold 2 predicted the 3D shapes of proteins to atomic accuracy, even where no similar structure was known. Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model (MuZero) arXiv preprint (published in Nature) · 19 Nov 2019MuZero matched AlphaZero at board games without being told the rules, by learning its own model of the game. Grandmaster level in StarCraft II using multi-agent reinforcement learning (AlphaStar) Nature · 2019AlphaStar reached Grandmaster rank in the online game StarCraft II playing all three of its races. Mastering Chess and Shogi by Self-Play with a General Reinforcement Learning Algorithm (AlphaZero) arXiv preprint · 5 Dec 2017One self-play learning method, starting from the rules alone, reached superhuman play in chess and shogi. Mastering the game of Go without human knowledge Nature · 19 Oct 2017AlphaGo Zero learned Go from self-play alone, with no human games, and beat the earlier champion-defeating AlphaGo 100 games to 0. Key facts and news (3) Models released Added this round: Gemma 2, CodeGemma, RecurrentGemma, PaliGemma, PaliGemma 2, ShieldGemma, ShieldGemma 2, DataGemma, Genie (first), Lyria (first), Lyria 2, Gemini Diffusion, Project Astra, RoboCat, Sparrow, RETRO, Perceiver/Perceiver IO, AlphaProof, AlphaGeometry 2, Gemini 2.0 Flash Thinking, Med-Gemini, Imagen 2, AlphaCode 2, Open X-Embodiment/RT-X, ALOHA Unleashed/DemoStart, Enformer, DGMR nowcasting, V2A, Veo (first), DeepNash, FermiNet, XLand agents, NotebookLM dialogue generation (BIZ-LAB-14001 to 14028, 14084 to 14088) one source News October 2026 "Google DeepMind and Singapore Government collaborate to build fine-tuned AI model for scam detection" — https://deepmind.google/blog/on-device-ai-scam-detection-mddi/ (post not opened; title from blog index) one source deepmind.google/discover/blog/ October 2026 "Accelerating the search for antimicrobial peptides with Google Antigravity" — https://deepmind.google/blog/accelerating-antimicrobial-discovery/ (post not opened; title from blog index) one source deepmind.google/discover/blog/ Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked.	London (EC2M 4AG)	Companies House · deepmind.google/blog/announcing-google-deepmind/ · deepmind.google/about · find-and-update.company-information.service.gov. · find-and-update.company-information.service.gov.	verified
GopherChips, models and products	A 280-billion-parameter language model that DeepMind used to study scale, risks and retrieval. released (paper)Key facts Announced 8 Dec 2021 with three papers (Gopher study; ethical and social risks; retrieval architecture). Figures (8 Dec 2021): models trained from 44 million to 280 billion parameters. Access: paper only. Successor named in Apr 2022: Chinchilla (70bn parameters).Made by Google DeepMind · 2021-12-08	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/language-modelling-at-scale · deepmind.google/blog/page/13	one source
IthacaChips, models and products	An AI model for ancient Greek inscriptions; the predecessor of Aeneas. released (paper)Key facts Blog March 2022 (index). Aeneas post (23 Jul 2025) says Ithaca is being upgraded to run on Aeneas.Made by Google DeepMind · 2022-03	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/12	one source
Juno-N-Coder-25B (Locai Labs)Chips, models and products	Locai Labs' compressed open-source coding model, cut down from NVIDIA Nemotron 3.5 Lightning to run on its Locai 1 office server. releasedKey facts 25.57B total parameters (19% fewer than the 31.56B base); experts pruned by 25%; quantised to NVFP4; listed as 25B-A3B; first Juno model shipped on the Locai 1 appliance (press release 11 Aug 2026)Made by Locai Labs · 2026-07	London	locailabs.com/research · locailabs.com/news/press-release-locai-labs-laun	one source
Jupiter-N (Locai Labs)Chips, models and products	Locai Labs' 120-billion-parameter model: NVIDIA's Nemotron 3 Super re-trained for British context; it powers Locai's GB1 assistant. releasedKey facts 120B parameters; post-train of NVIDIA Nemotron 3 Super; technical report arXiv 2604.17429; a sister 8B model, Jupiter-G, applies the same recipe to Google Gemma-4 (locailabs.com/research). Civo calls Jupiter the first model of the Mercury Series (14 Apr 2026)Made by Locai Labs · 2026-04	London	locailabs.com/research · civo.com/blog/uks-first-sovereign-llms-locai-lab	lab's own pages
Med-GeminiChips, models and products	Family of Gemini models specialised for medicine, tested on medical exam and imaging benchmarks. research (paper only)Key facts arXiv 2404.18416 submitted 29 Apr 2024; evaluated on 14 medical benchmarks; best model 91.1% accuracy (stated as state of the art); Google paper — split between Google Research and Google DeepMind authors not checked; access: paper onlyMade by Google DeepMind · 2024-04-29	London (EC2M 4AG) (via Google DeepMind)	arxiv.org/abs/2404.18416	one source
MedGemma (incl. MedGemma 1.5 4B)Chips, models and products	Open Gemma-based models for building health AI applications. released (open weights)Key facts Theme: General-purpose AI models. "MedGemma: Our most capable open models for health AI development" July 2025 (Google Research blog, listed on DeepMind index); MedGemma 1.5 4B January 2026 (Gemma page).Made by Google DeepMind · 2025-07	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/6/ · deepmind.google/models/gemma/	lab's own pages
Microsoft AI London hubCompanies · also on Big tech in the UK	A Microsoft research and engineering team in London that works on large language models and the infrastructure behind them. Profile Owner or parent Microsoft, Microsoft AI division (Mustafa Suleyman, EVP and CEO of Microsoft AI) Founded or opened announced 7 Apr 2024 Lab director or lead Jordan Hoffmann — hub lead (formerly Inflection and DeepMind) — from the 7 Apr 2024 announcement (not re-confirmed in 2026) Other leaders Mustafa Suleyman — EVP and CEO, Microsoft AI; leads the MAI Superintelligence Team (Nov 2025 post) Site and address as published "central London", with a team at Microsoft's Paddington office (Apr 2024) Headcount Not stated for the hub. Microsoft UK as a whole: "6,000 employees" — 16 Sep 2025 — https://blogs.microsoft.com/on-the-issues/2025/09/16/microsoft-30-billion-uk-ai-future/. On 31 Oct 2026 Microsoft AI listed 21 London roles (15 Microsoft AI, 6 Health) — https://microsoft.ai/job-region/london-united-kingdom/ What it does Microsoft AI's research and engineering hub in London. When it was announced in 2024, its stated job was work on state-of-the-art language models, their supporting infrastructure and tooling for foundation models. Its London job listings in October 2026 cover pretraining data, safety post-training, multimodal safety, frontier risk and alignment, AI supercomputing and health. Research groups or focus "advance state-of-the-art language models and their supporting infrastructure" (2024). London roles on 31 Oct 2026: Pretraining Data; Safety Post-training; Multimodal Safety; Production Safety; Research Scientist, Frontier Risk & Alignment, Futures team; Compute Infra HPC; AI Supercomputing; Health (Applied AI, Privacy, Product) Notable outputs (Microsoft AI division-wide; no page credits London) MAI-Thinking-1 — reasoning model, sparse Mixture of Experts, about 35B active and 1T total parameters, public preview on Microsoft Foundry — 12 Aug 2026; seven in-house MAI models (MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5, MAI-Image-2.5-Flash, MAI Transcribe-1.5, MAI-Voice-2, MAI-Voice-2-Flash) — 2 Jun 2026 Funding or budget as published Not stated for the hub. Microsoft's $30bn UK commitment for 2025–2028 is already DEAL-BIG-021 Partners named OpenAI — named as a collaborator of Microsoft AI London — Apr 2024; Surge — evaluation partner for MAI-Thinking-1 (division-wide) — Aug 2026 Uk government links HM Treasury release of 10 May 2024 cites the hub ("Microsoft recently announced a new AI hub in London"); AISI–Microsoft partnership of 5 May 2026 (Microsoft-wide; record BIZ-SAFE-2086) Status now Active: Microsoft AI listed 21 open London roles on 31 Oct 2026. Labs Round 8 deep profile (31 Oct 2026), as the researchers wrote it; confidence: verified (opening: Microsoft and GOV.UK); lead one source. Sources blogs.microsoft.com/blog/2024/04/07/announcing-n gov.uk/government/news/leading-tech-firms-invest microsoft.ai/job-region/london-united-kingdom/ microsoft.ai/news/introducing-mai-thinking-1/ microsoft.ai/news/building-a-hillclimbing-machin Why it is here: reason: hub set up to advance state-of-the-art language models and foundation-model tooling, led by Jordan Hoffmann (ex-Inflection, DeepMind); NOTE parent Microsoft AI's in-house MAI models (MAI-Thinking-1, MAI-Base-2) pages do not say where they were trained Key facts and news (3) Research focus Large language models and their supporting infrastructure; tooling for foundation models; works with other Microsoft AI teams and OpenAI; Paddington office blogs.microsoft.com/blog/2024/04/07/announcing-n News 15 October 2026 Microsoft AI completes pre-training of MAI-Base-2 (80bn active, 2.8tn total parameters, 16k GB200 GPUs over 38 days); training location not stated one source microsoft.ai/news/frontier-pre-training-at-micro 2 June 2026 Microsoft AI launches seven in-house MAI models (MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5, MAI-Image-2.5-Flash, MAI-Transcribe-1.5, MAI-Voice-2, MAI-Voice-2-Flash); page does not mention London one source microsoft.ai/news/building-a-hillclimbing-machin Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked.	London	blogs.microsoft.com/blog/2024/04/07/announcing-n · gov.uk/government/news/leading-tech-firms-invest · uktech.news/ai/microsoft-ai-hub-london-20240408 · microsoft.ai/news/frontier-pre-training-at-micro	verified
MuJoCo (open-sourced physics simulator)Chips, models and products	A physics simulator for robotics research that DeepMind opened up in 2021 and open-sourced in 2022. released (open source)Key facts "Opening up a physics simulator for robotics" October 2021; "Open-sourcing MuJoCo" May 2022 (index titles).Made by Google DeepMind · 2022-05	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/12 · deepmind.google/blog/page/13	one source
MuZeroChips, models and products	A game-playing system that learns its own model of a game and plans with it, without being told the rules. released (paper)Key facts Nature paper announced 23 Dec 2020 (preliminary paper 2019). Access: paper only. Figures (23 Dec 2020): playing strength rose by more than 1,000 Elo as thinking time per move rose from 0.1 to 50 seconds. Version history (as published): AlphaGo → AlphaZero → MuZero. Applied to video compression (index: "MuZero's first step from research into the real world", Feb 2022).Made by Google DeepMind · 2020-12-23Versions AlphaGo → AlphaZero → MuZeroRead the paper	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/muzero-mastering-go-chess-s · deepmind.google/blog/page/12	one source
Nemotron-Personas-UK-Extended (NVIDIA Nemotron Personas, en_GB)Chips, models and products	A synthetic dataset of UK population personas for training AI, built by NVIDIA with London-based Locai Labs. released (NVIDIA Dataset License)Key facts NGC catalogue: version 1.0 dated 20 Oct 2026; 1,000,000 records containing 9,000,000 personas; NVIDIA Dataset License Agreement (source data also under Open Government Licence v3.0 and CC BY 4.0). Extended version adds sensitive demographic fields and healthcare and finance personas not in the open version. Tile note: division and tile follow Locai Labs' Labs-page placement if different.Made by Locai Labs · 2026-10-20		catalog.ngc.nvidia.com/orgs/nvidia/nemotron-pers	one source
PaliGemmaChips, models and products	Small open model that reads images and text together, for tasks such as captioning and answering questions about pictures. released (open weights)Key facts Initial release 14 May 2024; 3B parameters; inspired by PaLI-3 (arXiv 2310.09199), uses SigLIP vision encoder and Gemma language model; access: open weightsMade by Google DeepMind · 2024-05-14Versions Gemma → Gemma 2 (and CodeGemma, RecurrentGemma, PaliGemma → PaliGemma 2, ShieldGemma → ShieldGemma 2, DataGemma)	London	ai.google.dev/gemma/docs/paligemma · ai.google.dev/gemma/docs/releases	lab's own pages
PaliGemma 2 (and PaliGemma 2 mix)Chips, models and products	Second version of the open image-and-text Gemma model, in three sizes. released (open weights)Key facts Released 5 Dec 2024 in 3B, 10B and 28B sizes; PaliGemma 2 mix released 19 Feb 2025 in 3B, 10B and 28B; access: open weightsMade by Google DeepMind · 2024-12-05Versions Gemma → Gemma 2 (and CodeGemma, RecurrentGemma, PaliGemma → PaliGemma 2, ShieldGemma → ShieldGemma 2, DataGemma)	London	ai.google.dev/gemma/docs/releases · ai.google.dev/gemma/docs/paligemma	lab's own pages
Perceiver and Perceiver IOChips, models and products	General neural-network designs that take in many kinds of data (images, sound, text) and produce many kinds of output. released (open code; papers)Key facts Blog 3 Aug 2021; papers arXiv 2103.03206 (Perceiver) and 2107.14795 (Perceiver IO); code open-sourced in deepmind-research GitHub repository; access: open codeMade by Google DeepMind · 2021-08-03	London	deepmind.google/blog/building-architectures-that	one source
Project GenieChips, models and products	An experimental app in which users create and explore AI-generated interactive worlds. released (prototype for Google AI Ultra subscribers in the US)Key facts Launched 29 Jan 2026 for Google AI Ultra subscribers in the US; Street View real-world places added 17 May 2026; Genie 4 coming to it (27 Oct 2026).Made by Google DeepMind · 2026-01-29	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rss.xml · deepmind.google/blog/rss.xml	lab's own pages
RecurrentGemma (2B and 9B)Chips, models and products	Open Gemma language model built on a different design (Griffin) that mixes recurrence with local attention to run more efficiently. released (open weights)Key facts Initial release 9 Apr 2024; 9B variant released 11 Jun 2024; based on the Griffin architecture (arXiv 2402.19427); access: open weightsMade by Google DeepMind · 2024-04-09Versions Gemma → Gemma 2 (and CodeGemma, RecurrentGemma, PaliGemma → PaliGemma 2, ShieldGemma → ShieldGemma 2, DataGemma)	London	ai.google.dev/gemma/docs/recurrentgemma · ai.google.dev/gemma/docs/releases	lab's own pages
Recursive FlintChips, models and products	Recursive Superintelligence's first public model: a 40-billion-parameter coding model built to run fast on local machines. releasedKey facts 40B total parameters, about 2B active, mixture-of-experts (recursive.com, read 2026-10-30); the company says its automated research system shaped every stage of the build; listed as "Model 01"Made by Recursive Superintelligence · 2026 (date not given on page)	London	recursive.com/ · recursive.com/	one source
RETRO (Retrieval-Enhanced Transformer)Chips, models and products	Language model that looks up passages in a very large text database while it writes, so a smaller model can match much larger ones. research (paper only)Key facts Announced 8 Dec 2021; 7.5B-parameter model with a retrieval database of up to 2 trillion tokens; outperformed 175B Jurassic-1 on 10 of 16 Pile datasets and 280B Gopher on 9 of 16; access: paper onlyMade by Google DeepMind · 2021-12-08	London	deepmind.google/blog/improving-language-models-b	one source
RT-2 (Robotic Transformer 2)Chips, models and products	A robot-control model that learns from web and robot data and turns camera images and instructions into robot actions. released (paper)Key facts Announced 28 Jul 2023 (paper only). Figures (28 Jul 2023): success on unseen scenarios rose from RT-1's 32% to 62%; over 6,000 robot trials. Predecessor: RT-1.Made by Google DeepMind · 2023-07-28	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/rt-2-new-model-translates-v · deepmind.google/blog/page/9/	one source
Stability AICompanies	Builds open-weight generative models that make images, video and audio from text, including Stable Diffusion and Stable Audio, for media, games and music companies. Profile Founders null (not named on pages opened) Chief executive Prem Akkaraju — CEO from 25 Jun 2024 (former CEO of Weta Digital) Chief scientist or research lead Christian Laforte — CTO (June 2024; previously interim co-CEO from March 2024) Other leaders Sean Parker — Executive Chairman (from 25 Jun 2024); Shan Shan Wong — COO (June 2024); Robert Legato — Chief Pipeline Architect (19 Mar 2025); Thomas Laffont (Coatue co-founder) — joined board 25 Aug 2026; Dana Settle (Greycroft) — board; James Cameron — board Hq null (2024 and 2026 releases datelined Los Angeles; no HQ stated). Site already holds a 2026 UK office closure (DEAL-LAB-014 / DEAL-LAB-2026). What it does Generative AI company that makes open-weight models for images (Stable Diffusion), audio and music (Stable Audio), video, 3D and multi-view camera work, plus enterprise creative tools such as Brand Studio. It now works closely with record labels, a games publisher and an advertising group. Research focus open-weight image, video, 3D and audio generation; enterprise creative production Models and products Stable Audio 3.0 — open-weight music model family — 20 May 2026; Brand Studio — creative production platform — 8 Apr 2026; Stable Audio 2.5 — enterprise sound production — 10 Sep 2025; Stable Diffusion 3.5 NIM (with NVIDIA) — 12 Aug 2025; Stability AI Solutions — 5 Aug 2025; Stable Video 4D 2.0 — 20 May 2025; Stable Audio Open Small (with Arm) — 14 May 2025; Stable Virtual Camera — multi-view video — 18 Mar 2025 Customers or partners named Warner Music Group — AI music creation tools — 19 Nov 2025 — https://stability.ai/news-updates; Universal Music Group — strategic alliance on professional AI music tools — 30 Oct 2025 — same url; Electronic Arts — game development partnership — 23 Oct 2025 — same url; Amazon Bedrock — image services — 18 Sep 2025 — same url; NVIDIA — SD 3.5 NIM — 12 Aug 2025 — same url; Arm — 14 May 2025 — same url; AMD — 16 Apr 2025 — same url; joined the Tech Coalition — 11 Feb 2026 — same url Funding total as published 232,000,000 USD, "inclusive of two equity rounds and convertible notes" — as of 25 Aug 2026 — https://stability.ai/news-updates/stability-ai-latest-funding-backed-by-entertainment-industry-biggest-names Rounds found 2024-06-25 — initial round under new management — undisclosed — no lead named — Greycroft, Coatue, Sound Ventures, Lightspeed, O'Shaughnessy Ventures, Sean Parker, Prem Akkaraju, Eric Schmidt, Robert Nelsen; 2026-08-25 — Series B — 76,000,000 USD new capital — no lead named — Electronic Arts, Sony Music Group, Universal Music Group, Warner Music Group, AMD Ventures, Pacific Alliance Ventures (new); Coatue, Greycroft, Kadmos Capital, Lightspeed, Mantis Capital, Sound Ventures, WPP (returning); individuals incl. Sean Parker, Eric Schmidt, James Cameron Status now Raised a $76m Series B in August 2026 backed by music labels, EA and Sony Music; releases are datelined Los Angeles. Labs Round 7 deep profile (30 Oct 2026), as the researchers wrote it; confidence: single_source. Sources stability.ai/news-updates/stability-ai-latest-fu stability.ai/news-updates/stability-ai-latest-fu stability.ai/news-updates/stability-ai-secures-s stability.ai/news-updates Models and products (27) Version chainsStable Audio 2.0 (3 Apr 2024) → Stable Audio Open (5 Jun 2024) → Stable Audio Open Small (14 May 2025) → Stable Audio 2.5 (10 Sep 2025) → Stable Audio 3.0 (20 May 2026)StableLM → Stable LM 2Stable Diffusion v1 → 2.0 → SDXL → SDXL Turbo → SD3 → SD3.5 Stable Audio 3.0 2026-05-20 released (open weights for Small SFX, Small, Medium; Large via API)Stability AI's audio model family for music and sound effects, trained on licensed data, mostly with open weights. Brand Studio 2026-04-08 released (product)Stability AI's creative production platform that generates content in a company's own brand style. Stable Audio 2.5 2025-09-10 released (API and enterprise licence)An enterprise audio model that generates brand music and sound in seconds. Stability AI Solutions 2025-08-05 released (enterprise service)Stability AI's packaged generative AI services for enterprise creative production. Stable Audio Open Small 2025-05-14 released (open weights)A small open audio model, released with Arm, that generates sound on phones and other devices. Stable Virtual Camera 2025-03-18 released (model)A model that makes multi-view video from images with 3D camera control. Stable Point Aware 3D (SPAR3D) 2025-01-08 released (model)A model that generates a complete 3D object from an image and allows real-time editing. Stable Diffusion 3.5 2024-10-22 released (model)Stability AI's text-to-image model family released in October 2024 and later offered through cloud marketplaces. Stable Fast 3D 2024-08-01 released (model)A model that makes a 3D asset from a single image quickly. Stable Video 4D (and SV4D 2.0) 2024-07-24 released (model)A model that turns one video into multi-angle video of a moving object. Stable Assistant 2024-07-09 released (product)Stability AI's assistant product, through which its image models were offered to users. Stable Audio Open 2024-06-05 released (open source)An open-source model for making short audio samples and sound effects. Stable Audio 2.0 2024-04-03 released (product)Stability AI's second-generation music and sound generation model. Stable Video 3D 2024-03-18 released (model)A model that generates new viewing angles and 3D objects from one image. TripoSR 2024-03-05 released (model)A model that makes a 3D object from a single image quickly. Stable Diffusion 3 (incl. SD3 Medium) 2024-02-22 released (API; open weights for Medium)Stability AI's third-generation text-to-image model; the 2-billion-parameter Medium version was released with open weights. Stable Cascade 2024-02-12 released (model)A Stability AI model announced in February 2024; launch post not opened, so details are not recorded. Stable LM 2 (1.6B and 12B) 2024-01-19 released (model)Stability AI's second family of small open language models, in 1.6-billion and 12-billion-parameter sizes. Stable Code 3B 2024-01-16 released (model)A small (3-billion-parameter, per its name) coding language model from Stability AI; launch post titled "Coding on the Edge". Stable Zero123 (and Stable Zero123C) 13 Dec 2023 (post updated 8 Jan 2024) released (non-commercial; Zero123C commercial for members)Model that generates new views of an object from a single picture, as a step to making 3D objects. SDXL Turbo (Adversarial Diffusion Distillation) 2023-11-28 released (non-commercial open weights)Fast version of SDXL that makes an image in a single step, close to real time. Stable Video Diffusion 2023-11-21 research preview (open code and weights)Stability AI's first video model, which turns a single image into a short video clip. SDXL (Stable Diffusion XL) 2023-07-04 (paper) superseded (open weights)Larger Stable Diffusion model for higher-resolution images, with a second text encoder and an optional refining model. DeepFloyd IF 28 Apr 2023 (year from page image timestamps) released (research licence)Text-to-image model from Stability AI's DeepFloyd research lab, built in three stages and able to write legible text inside images. Stable Diffusion (first public release, v1) no date superseded (open weights)The original open text-to-image model that made pictures from typed prompts on an ordinary consumer graphics card. Stable Diffusion 2.0 no date superseded (open weights)Second major Stable Diffusion release, with a new text encoder, higher default resolution and a 4x upscaler. StableLM (Alpha, 3B and 7B) no date superseded (open weights)Stability AI's first suite of open language models, released in two small sizes. Why it is here: reason: trains its own language models (Stable LM 2 1.6B technical report, Feb 2024) Selected papers (12) Arbor: Explicit Geometric Conditioning for Controllable 3D Asset Generation Stability AI research · 7 October 2026 4Director: Controlling Video World Models with Rigid 3D Geometry 7 October 2026 SemanTok: Predictable Semantic Tokens for Efficient Autoregressive Video Generation 7 October 2026 Stable-Layers 3 June 2026 Stable Audio 3 and SAME 20 May 2026 Stable Audio 3 arXiv preprint · 18 May 2026The Stable Audio 3 model family generates music and sound in under two seconds on one H200 GPU and can edit or extend recordings. SD3.5-Flash: Distribution-Guided Distillation of Generative Flows 26 September 2025 Stable Audio Open arXiv preprint · 19 Jul 2024An open-weight text-to-audio model that performed on a par with the best systems on several measures. Scaling Rectified Flow Transformers for High-Resolution Image Synthesis (Stable Diffusion 3 paper) arXiv preprint · 5 Mar 2024The architecture behind Stable Diffusion 3; the biggest models outperformed the best image generators of the time. Adversarial Diffusion Distillation (SDXL Turbo method) arXiv preprint · 28 Nov 2023A training method that lets a large image model produce a picture in a single step, in real time. Stable Video Diffusion: Scaling Latent Video Diffusion Models to Large Datasets arXiv preprint · 25 Nov 2023Training on large video datasets produced a text-to-video model competitive with closed-source systems. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis arXiv preprint · 4 Jul 2023The SDXL image model clearly outperformed earlier versions of Stable Diffusion; code and weights released on Stability AI's GitHub. Key facts and news (7) Status now Led by CEO Prem Akkaraju (appointed June 2024); $76m new capital in Series B (25 Aug 2026) brought total funding to $232m; Coatue co-founder Thomas Laffont joined the board; announcement datelined Los Angeles one source stability.ai/news-updates/stability-ai-latest-fu Funding as published (reported) $232m including two equity rounds and convertible notes (25 Aug 2026) one source stability.ai/news-updates/stability-ai-latest-fu Users and customers Partners and backers named: Electronic Arts, Sony Music Group, Universal Music Group, Warner Music Group, WPP, AWS (Amazon Bedrock), NVIDIA, AMD, Arm stability.ai/news Research focus 3D asset generation, controllable video world models, efficient autoregressive video, image layer decomposition, KV-cache compression, music autoencoders and text-to-audio (2025–26 research posts) stability.ai/research Models released Core Models list (updated 20 May 2026) also names SD Turbo, Japanese SDXL, Japanese Stable CLIP, Stable LM Zephyr 3B, Japanese Stable LM 2 1.6B, Japanese Stable VLM and Stable Code Instruct 3B (no dates on page; not added as products — UK build not shown for the Japanese models) one source stability.ai/core-models News 18 August 2026 Stable Audio plugin for digital audio workstations and new StableAudio.com experience (beta), up to six-minute generations one source stability.ai/news-updates/sharing-a-new-way-to-w 11 February 2026 Stability AI joins the Tech Coalition one source stability.ai/news-updates/stability-ai-joins-the Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked.	Los Angeles	Companies House · stability.ai/news-updates/stability-ai-latest-fu · tech.eu/2026/08/25/stabiityai-raises-76m-with-wa · stability.ai/news/stability-ai-secures-significa · find-and-update.company-information.service.gov.	verified
Stable LM 2 (1.6B and 12B)Chips, models and products	Stability AI's second family of small open language models, in 1.6-billion and 12-billion-parameter sizes. released (model)Key facts Theme: General-purpose AI models. Stable LM 2 1.6B announced 19 Jan 2024; Stable LM 2 12B announced 8 Apr 2024.Made by Stability AI · 2024-01-19Versions StableLM → Stable LM 2	Los Angeles (via Stability AI)	stability.ai/news-updates/category/Product · stability.ai/news-updates/category/Product	one source
StableLM (Alpha, 3B and 7B)Chips, models and products	Stability AI's first suite of open language models, released in two small sizes. superseded (open weights)Key facts Alpha models of 3 billion and 7 billion parameters, with 15 to 65 billion stated to follow; trained on 1.5 trillion tokens of content; base models CC BY-SA-4.0, fine-tuned research models CC BY-NC-SA 4.0; access: open weights (GitHub); successor Stable LM 2 already listed (BIZ-LAB-7039)Made by Stability AIVersions StableLM → Stable LM 2	London	stability.ai/news-updates/stability-ai-launches-	one source
Streams (DeepMind Health app)Chips, models and products	A mobile app for NHS hospital clinicians; DeepMind published a post explaining why it did not use AI. released (NHS hospital app)Key facts Index posts: "Bringing Streams to Yeovil District Hospital NHS Foundation Trust" (Nov 2017); "Why doesn't Streams use AI?" (Nov 2017); Royal Free legal audit findings (Jun 2018); "Scaling Streams with Google" (Nov 2018).Made by Google DeepMind · 2017-11	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/17/ · deepmind.google/blog/page/16	one source
SynthIDChips, models and products	An invisible watermark that marks AI-generated images, then text, video and audio, so they can be identified later. released (in Google products; text watermarking open-source planned May 2024)Key facts Launched 29 Aug 2023 as a limited beta for Vertex AI customers using Imagen. Extended 14 May 2024 to text in the Gemini app and to Veo video; open-sourcing of text watermarking announced as planned. Later: SynthID Detector portal (May 2025, index); SynthID Bio for proteins (30 Sep 2026). Veo and Lyria outputs carry SynthID (model pages).Made by Google DeepMind · 2023-08-29	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/identifying-ai-generated-im · deepmind.google/blog/watermarking-ai-generated-t	lab's own pages
T5Gemma (and T5Gemma 2)Chips, models and products	A collection of Gemma models with an encoder-decoder design. released (open weights)Key facts Theme: General-purpose AI models. T5Gemma July 2025 ("A new collection of encoder-decoder Gemma models"); T5Gemma 2 December 2025 (Gemma page).Made by Google DeepMind · 2025-07	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/6/ · deepmind.google/models/gemma/	lab's own pages
VaultGemmaChips, models and products	A Gemma language model trained with differential privacy, a privacy-protecting training method. released (open weights)Key facts Theme: General-purpose AI models. September 2025 (Google Research blog listed on DeepMind index; also on Gemma page).Made by Google DeepMind · 2025-09	London (EC2M 4AG) (via Google DeepMind)	deepmind.google/blog/page/5/ · deepmind.google/models/gemma/	lab's own pages