| ATUM (Lloyds Banking Group transaction foundation model)Chips, models and products | Lloyds Banking Group's in-house foundation model that turns billions of banking transactions into reusable AI. builtKey facts No size or date published on the page read; NVIDIA lists it among UK deployments on 13 Oct 2026Made by Lloyds Banking Group AI programme · 2026 (reported 2026-10-13) | Edinburgh (registered office) (EH1 1YZ) (via Lloyds Banking Group AI programme) | blogs.nvidia.com/blog/uk-ecosystem-deployment/ · cic.vc/nvidia-announces-new-investments-in-the-u | one source |
| CaliBenchChips, models and products | Odyssey benchmark that tests whether world models are physically calibrated. released (benchmark)Key facts Announced 10 Aug 2026; post not opened (link blocked)Made by Odyssey · 2026-08-10 | | odyssey.systems/writing | one source |
| CosineCompanies · also on Software and tools | Makes Genie, an AI software engineer that fixes bugs, builds features and rewrites code on its own or with human developers.coding agentsProfile- Founded
- 2022 (UKTN)
- Founders
- Alistair Pullen — co-founder and CEO; Yang Li — co-founder, COO; Sam Stenner — co-founder (role not stated)
- Chief executive
- Alistair Pullen — from null
- Hq
- null (UKTN does not state)
- What it does
- Builds Genie, an AI coding agent ("Artificial Developer") that fixes bugs, builds features and refactors code alone or alongside human developers. Backed by Y Combinator.
- Products
- Genie — 2024-08
- Usage or revenue as published
- 30% on the SWE-Bench coding benchmark (company claim) — 2024-08-13 — UKTN
- Rounds found
- 2024-08-13 — round — 2.5m USD (1.9m GBP) — Uphonest and SOMA Capital — Lakestar, Focal and others (already DEAL-SW-019; Lakestar and Focal new detail)
- Status now
- Active per existing record; no 2025–2026 page opened this round.
- Confidence
- single_source
Round 6 deep profile (30 Oct 2026), as the researchers wrote it; sources are listed in the record panel. Models and products (3)- Lumen Sovereign (Cosine) 2026-06-08 in training (target late 2026; commercial launch Q1 2027)Cosine's planned UK-trained frontier model, co-designed with 13 British banks, defence firms and others, to run inside customers' own systems.
- Lumen models (Lumen Sovereign 256K, Lumen Scout, Lumen Outpost) 2026 in development (alpha planned 2026 Q4; commercial launch planned 2027 Q1)Cosine's own model family; Lumen Sovereign is a planned frontier model that organisations can run on their own infrastructure, even air-gapped.
- Cosine coding agent (CLI, Desktop and Cloud) no date released (product)Cosine's autonomous AI software engineer that works alongside human developers through a command line, desktop app and cloud.
| London (and San Francisco) | tech.eu/2024/08/12/uk-ai-startup-cosine-raises-2 · gov.uk/government/news/ai-firms-pioneering-drug- · tech.eu/2026/04/17/cosine-goes-from-benchmark-le · ycombinator.com/companies/cosine | one source |
| Cosine coding agent (CLI, Desktop and Cloud)Chips, models and products | Cosine's autonomous AI software engineer that works alongside human developers through a command line, desktop app and cloud. released (product)Key facts About page (read 30 Oct 2026): surfaces CLI, Desktop, Cloud; install via Homebrew; 2024 timeline entry: "achieved a historic 30% on SWE-bench" as alpha design partners for GPT-4 Turbo. Product name "Genie" not shown on the pages opened. No launch date published there.Made by Cosine | London (and San Francisco) (via Cosine) | cosine.sh/about | one source |
| Literal LabsCompanies · also on Chips and hardware | Builds a form of AI based on logic rules rather than neural networks, designed to run on very little power on small chips.Profile- Founded
- null (spun out of Newcastle University)
- Founders
- Alex Yakovlev — academic co-founder (spun the company out); Rishad Shafik — academic co-founder (spun the company out)
- Chief executive
- Noel Hurley — from 2023 (former Arm VP)
- Other leaders
- Leon Fedden — chief technology officer (former AI deep learning platform lead, AstraZeneca)
- Hq
- Newcastle (England: North East); other_offices: null
- Headcount
- null ("has since doubled its headcount", June 2025)
- What it does
- A Newcastle University spin-out building logic-based AI models that it says are faster and use less energy than neural networks. In June 2025 it said its first commercial product was due later that year.
- Research focus
- logic-based AI; energy-efficient models
- Models and products
- first commercial product announced as due later in 2025 — June 2025 (name null)
- Funding total as published
- 4,600,000 GBP (pre-seed) — as of 2 June 2025 — https://www.uktech.news/ai/ai-firm-from-ex-arm-exec-scores-4-6m-investment-20250602
- Rounds found
- 2025-06-02 — pre-seed — 4,600,000 GBP — Northern Gritstone (lead), Mercuri (co-lead) — Sure Valley Ventures, Cambridge Future Tech SPV, angels (already DEAL-CHIP-010 / DEAL-LAB-2018)
- Status now
- null — no news after June 2025 found in this round (the deal file's 13 November 2025 "rename" entry was not checked).
- Legal name
- Not stated (Companies House number 14746541 shown on site — lead to check)
- Founded (Labs Round 8)
- 2023 (spun out of Newcastle University)
- Founders (Labs Round 8)
- Alex Yakovlev — co-founder; Rishad Shafik — co-founder (company builder: Cambridge Future Tech)
- Chief executive (Labs Round 8)
- Noel Hurley — from 2023 (former Arm executive)
- Chief scientist or research lead
- Leon Fedden — CTO
- Other leaders (Labs Round 8)
- Mike Park — CFO; Daniel P Dykes — Head of Product; Jim Darragh — Non-Executive Chairman (from 4 Mar 2026); Jem Davies — non-executive director; Ole-Christoffer Granmo — chair, Technical Steering Committee
- Hq (Labs Round 8)
- Newcastle upon Tyne (England: North East); other_offices: not stated
- What it does (Labs Round 8)
- Literal Labs builds AI models from logic rather than neural networks ("Logic-Based Networks", from Tsetlin machine research). It says they run faster and with far less energy on ordinary processors, without GPUs, and give the same output for the same input. ModelMill is its platform for training and deploying them.
- Research focus (Labs Round 8)
- Tsetlin machines and logic-based AI for edge devices; explainable, energy-efficient inference.
- Models and products (Labs Round 8)
- ModelMill — platform to train and deploy Logic-Based Networks — not stated; Logic-Based Networks (LBNs) — logic-based AI models, claimed up to 50x faster inference and 50x less energy than neural networks — not stated
- Funding total as published (Labs Round 8)
- £4.6m — as of 2 Jun 2025 — https://www.literal-labs.ai/press/literal-labs-pre-seed-round.html
- Rounds found (Labs Round 8)
- 2 Jun 2025 — pre-seed — £4.6m — Northern Gritstone and Mercuri (co-leads) — Sure Valley Ventures, Cambridge Future Tech SPV, angels (already DEAL-LAB-2018 / DEAL-CHIP-010)
- Awards and programmes
- Global AI Excellence Awards, AI Research & Innovation prize — Mar 2026; National AI Awards 2025 AI Innovation shortlist — 2025
- Status now (Labs Round 8)
- Trading; appointed Jim Darragh as non-executive chairman in March 2026.
Labs Round 7 and Labs Round 8 deep profile (30–31 Oct 2026), as the researchers wrote it; confidence: verified (funding: company plus trade); people one source. Sources - uktech.news/ai/ai-firm-from-ex-arm-exec-scores-4
- literal-labs.ai/about/
- literal-labs.ai/press/literal-labs-pre-seed-roun
- tech.eu/2025/06/02/literal-labs-raises-ps46m-to-
- literal-labs.ai/about/news/jim-darragh-appointed
| Newcastle upon Tyne (NE1 5JE) | Companies House · find-and-update.company-information.service.gov. · literal-labs.ai/press/literal-labs-pre-seed-roun · literal-labs.ai/about/news/ · find-and-update.company-information.service.gov. | verified |
| Locai LabsCompanies | UK sovereign AI company; released Nemotron-Personas-UK, an open dataset of one million synthetic personas, with NVIDIA.Profile- What it does
- Described by NVIDIA as a "sovereign AI company". In October 2026 it released Nemotron-Personas-UK, an open-source dataset of one million synthetic personas that reflect the UK's demographic, regional and cultural mix, to help developers build AI that better represents UK users.
- Models and products
- Nemotron-Personas-UK — open dataset of one million synthetic UK personas — October 2026
- Customers or partners named
- NVIDIA — Nemotron-Personas-UK release named in NVIDIA's UK post — 13 October 2026 — https://blogs.nvidia.com/blog/uk-ecosystem-deployment/
- Rounds found
- null (the £1m Fuel Ventures pre-seed of 22 April 2026, DEAL-SW-050, belongs to a different "Locai", not Locai Labs; see Build notes)
- Status now
- Active; released Nemotron-Personas-UK with NVIDIA in October 2026.
Labs Round 7 deep profile (30 Oct 2026), as the researchers wrote it; confidence: verified (dataset release only). Sources - blogs.nvidia.com/blog/uk-ecosystem-deployment/
- itbrief.co.uk/story/nvidia-backs-five-uk-startup
Models and products (3)- Nemotron-Personas-UK-Extended (NVIDIA Nemotron Personas, en_GB) 2026-10-20 released (NVIDIA Dataset License)A synthetic dataset of UK population personas for training AI, built by NVIDIA with London-based Locai Labs.
- Juno-N-Coder-25B (Locai Labs) 2026-07 releasedLocai Labs' compressed open-source coding model, cut down from NVIDIA Nemotron 3.5 Lightning to run on its Locai 1 office server.
- Jupiter-N (Locai Labs) 2026-04 releasedLocai Labs' 120-billion-parameter model: NVIDIA's Nemotron 3 Super re-trained for British context; it powers Locai's GB1 assistant.
| | blogs.nvidia.com/blog/uk-ecosystem-deployment/ · locailabs.com/ · locailabs.com/about · civo.com/newsroom/locai-labs-and-civo-unveil-pro | one source |
| Lumen models (Lumen Sovereign 256K, Lumen Scout, Lumen Outpost)Chips, models and products | Cosine's own model family; Lumen Sovereign is a planned frontier model that organisations can run on their own infrastructure, even air-gapped. in development (alpha planned 2026 Q4; commercial launch planned 2027 Q1)Key facts Lumen Sovereign page (read 30 Oct 2026): development 2026 Q2; alpha for coalition members 2026 Q4; commercial launch 2027 Q1 on cloud, VPC or air-gapped GPUs; "a 1T parameter intermediate checkpoint with 49.2B active parameters"; lead configuration expected 192–256 experts, top-6 routing; Vodafone among coalition partner logos (Vodafone memorandum already held as BIZ-CLD-025). About page: "Shortly after releasing our first Lumen models" (2026). No benchmarks published.Made by Cosine · 2026 | London (and San Francisco) (via Cosine) | cosine.sh/lumen-sovereign · cosine.sh/about | lab's own pages |
| Lumen Sovereign (Cosine)Chips, models and products | Cosine's planned UK-trained frontier model, co-designed with 13 British banks, defence firms and others, to run inside customers' own systems. in training (target late 2026; commercial launch Q1 2027)Key facts Target: 1T-parameter intermediate checkpoint with 49.2B active parameters, 256k-context sparse mixture-of-experts; 500,000 GH200 GPU-hour budget on Isambard-AI via the UK Sovereign AI programme; about one-third of compute for software engineering, two-thirds for areas such as cyber testing and KYC/AML; coalition of 13 (BAE Systems, Babcock, BT, Era4, HSBC, Leonardo, Lloyds Banking Group, LSEG, NatWest Group, PwC, Telefónica Tech UK&I, Thales UK, The Alan Turing Institute) announced 8 Jun 2026Made by Cosine · 2026-06-08 | London | cosine.sh/lumen-sovereign · cosine.sh/blog/building-lumen-sovereign-uk-indus | lab's own pages |
| MachentaCompanies · also on Start-ups | Builds an AI model that lets robots learn new physical tasks, such as packing and assembly, after one demonstration.Profile- Founded
- Not stated (work began about six years before Oct 2026 at Imperial College London)
- Founders
- Edward Johns — co-founder (Professor of robot learning, Imperial College London); Vitalis Vosylius — co-founder (PhD student of Johns)
- Hq
- London, King's Cross (England: London); other_offices: not stated
- Headcount
- around a dozen — 27 Oct 2026 — https://tech.eu/2026/10/27/london-robotics-startup-machenta-raises-40m-seed-investment/ (single_source)
- What it does
- Machenta builds an AI model that lets robots learn a new physical task, such as packing, assembling or tidying, after seeing it done once. The model is trained in simulation rather than on real-world robot data.
- Research focus
- One-shot robot learning trained in simulation (Johns likens it to "physical vibe coding").
- Models and products
- robot-learning foundation model (unnamed) — not stated
- Funding total as published
- $40m — as of 27 Oct 2026 — https://tech.eu/2026/10/27/london-robotics-startup-machenta-raises-40m-seed-investment/
- Rounds found
- 27 Oct 2026 — seed — $40m — Creandum and Union Square Ventures (co-leads) — not stated (already DEAL-LAB-2010 / DEAL-LC-139)
- Status now
- Emerged from stealth on 27 October 2026 with a $40m seed round, described as the biggest seed round raised by a UK robotics company.
Labs Round 8 deep profile (31 Oct 2026), as the researchers wrote it; confidence: one source. Sources - tech.eu/2026/10/27/london-robotics-startup-mache
| London (N19 4AJ ) | Companies House · tech.eu/2026/10/27/london-robotics-startup-mache · find-and-update.company-information.service.gov. · find-and-update.company-information.service.gov. | verified |
| Malted AICompanies · also on Software and tools | Makes small, task-specific AI language models for businesses by distilling the outputs of large models.developer toolsProfile- Founded
- 2023 (UKTN)
- Founders
- Iain Mackie — co-founder and CEO
- Chief executive
- Iain Mackie — from 2023
- Hq
- Edinburgh (Scotland)
- What it does
- A "distillation" platform that turns the output of large language models into smaller, tailored AI models for a firm's own tasks, which it says cost less to run.
- Products
- distillation platform for small, task-specific language models — 2024
- Customers named
- null ("piloting its small language models with financial and legal firms", 2024-05)
- Rounds found
- 2023 — Pre-seed — 1m GBP — Creator Fund — null; 2024-05-15 — Seed — 6m GBP — Hoxton Ventures — Creator Fund, angels (already DEAL-LB-124 / DEAL-RG-001)
- Status now
- Active per existing record; no 2025–2026 page opened this round.
- Confidence
- single_source
- Founded (Labs Round 8)
- 2023 (founders met at the University of Glasgow in 2019)
- Founders (Labs Round 8)
- Iain Mackie — founder and CEO; Carlos Gemmell — founder and Chief Scientist; Federico Rossetto — founder and Chief Architect
- Chief scientist or research lead
- Carlos Gemmell — Chief Scientist
- Other leaders
- Ewen Fleming — Chief Commercial Officer; Sandy Nairn — Chairman; Matt Davies — Head of Engineering; Rachel Young — Head of Operations; Jack Merriman — Head of Forward Deployment; Jim Croft — Head of SRE
- What it does (Labs Round 8)
- Malted builds small, specialised language models by distillation for financial services. Its platform analyses every customer call, chat and email at banks, building societies and wealth managers to find signals about customer outcomes and to give evidence to regulators (Consumer Duty).
- Research focus
- Small language models for specific tasks and proprietary distillation methods; information retrieval, conversational AI, reasoning with external tools.
- Models and products
- Malted platform — analysis of 100% of customer interactions — not stated; Pulse — product named in a customer testimonial — not stated; specialised small language models — built by distillation — not stated
- Customers or partners named
- Scottish Building Society — customer testimonial (Paul Denton, CEO) — not stated — https://malted.ai/; Skipton Building Society — customer testimonial — not stated — https://malted.ai/; Openwork — customer testimonial (Rob Barker, CEO) — not stated — https://malted.ai/
- Funding total as published
- Not stated (company page: £1m raised in 2023, £6m in 2024)
- Rounds found (Labs Round 8)
- 2023 — first round — £1m — not stated — not stated; 15 May 2024 — seed — £6m — Hoxton Ventures (already DEAL-LB-124 / DEAL-RG-001 / DEAL-LAB-2016)
- Awards and programmes
- Sequoia F100 Rising Star (date not stated); Tech Nation Future Fifty (date not stated); founders' Amazon Alexa Prize challenge win (2021)
- Status now (Labs Round 8)
- Trading; sells specialised AI to UK banks and building societies, with named customers.
Labs Round 7 and Labs Round 8 deep profile (30–31 Oct 2026), as the researchers wrote it; confidence: one source. Sources - malted.ai/company
- malted.ai/
Models and products (1)- Malted Platform (specialised small language models) — (launch date not stated) live (product)Malted AI's system that uses small, specialised language models to analyse customer calls, chats and emails for financial services.
| Edinburgh | Companies House · uktech.news/ai/malted-ai-funding-20240515 · find-and-update.company-information.service.gov. · find-and-update.company-information.service.gov. · technation.io/unveils-latest-cohort-of-tech-scal | verified |
| Malted Platform (specialised small language models)Chips, models and products | Malted AI's system that uses small, specialised language models to analyse customer calls, chats and emails for financial services. live (product)Key facts Company states its distilled small models are "100x more efficient than large general-purpose models"; analyses 100% of customer interactions in real time; deployed in single-tenant environments with regional data residency; access: productMade by Malted AI · — (launch date not stated) | Edinburgh (via Malted AI) | malted.ai/ | one source |
| OdysseyCompanies | World-model company founded by British researchers, with California operations and a UK research team; given UK AIRR compute.Profile- Legal name
- null ("Odyssey ML" in The Decoder)
- Founded
- 2023
- Founders
- Oliver Cameron — co-founder and CEO (former co-founder and CEO of Voyage; later VP of product at Cruise); Jeff Hawke — co-founder and CTO (former Wayve engineer)
- Chief executive
- Oliver Cameron — from null
- Chief scientist or research lead
- Jeff Hawke — CTO
- Hq
- Palo Alto, California (AI Business) / Menlo Park (Tech Funding News) (Outside the UK); other_offices: London; Zurich (The Decoder: team "works out of London, Zurich, and Palo Alto")
- Headcount
- 55 — 17 June 2026 — https://the-decoder.com/amazon-nvidia-and-amd-bet-310-million-on-ai-startup-building-3d-world-models/ (trade)
- What it does
- A lab founded by two British AI researchers that builds "world models": AI that learns from video, sound and text to produce interactive, real-time simulated worlds. It is known for turning text prompts into interactive video. Its models are aimed at games, robotics, simulation and defence. It keeps a UK research team alongside its California base.
- Research focus
- world models; multimodal learning (vision, language, sound); interactive video; multi-agent shared worlds
- Models and products
- Odyssey-2 Max — world model (scale and physics) — date null; Starchild-1 — real-time multimodal audio and video — date null; Agora-1 — multi-agent shared world state — shown live in London 12 June 2026; PROWL — reinforcement learning agent that probes the world model — date null
- Customers or partners named
- AWS — preferred cloud provider; models to be optimised for Trainium chips — 17 June 2026 — https://techcrunch.com/2026/06/17/world-model-maker-odyssey-nabs-1-45b-valuation-backed-by-amazon-and-other-big-names/
- Compute
- AWS preferred cloud; Trainium chips (June 2026); Sovereign AI / AIRR compute in the UK (April 2026)
- Funding total as published
- 337,000,000 USD — as of 17 June 2026 — https://techcrunch.com/2026/06/17/world-model-maker-odyssey-nabs-1-45b-valuation-backed-by-amazon-and-other-big-names/
- Valuation reported
- 1,450,000,000 USD — June 2026 — TechCrunch; Air Street Press (investor); AI Business; The Decoder
- Rounds found
- late 2023 — seed — amount null — Air Street Capital wrote "the first check" — null; 2024-11-13 — Series A — 18,000,000 USD — EQT Ventures — null (NEW: DEAL-LAB-9002); 2026-06-17 — Series B — 310,000,000 USD — Natural Capital — Amazon, AMD Ventures, GV, EQT, IQT; angels Jeff Dean, Elad Gil, Garry Tan, Guillermo Rauch, Kyle Vogt (NEW: DEAL-LAB-9001)
- Uk government links
- Sovereign AI / AIRR compute — 16 April 2026 (already DEAL-LAB-2035)
- Status now
- Active; raised a $310m Series B at a $1.45bn valuation in June 2026, with teams in London, Zurich and Palo Alto.
Labs Round 7 deep profile (30 Oct 2026), as the researchers wrote it; confidence: verified (Series B, valuation); Series A and headcount single_source. Sources - press.airstreet.com/p/odyssey-series-b
- techcrunch.com/2026/06/17/world-model-maker-odys
- the-decoder.com/amazon-nvidia-and-amd-bet-310-mi
- gov.uk/government/news/ai-firms-pioneering-drug-
- maginative.com/odyssey-raises-18m-to-build-ai-wo
- aibusiness.com/generative-ai/world-model-ai-lab-
Models and products (11)Version chainsOdyssey-1 → Odyssey-2 (Pro, Max) → Odyssey-3 - Agora-2 2026-09-24 announcedSecond version of Odyssey's multi-agent world simulation model.
- Odyssey-3 2026-09-15 announced (Sep 2026); further release post Oct 2026Odyssey's third-generation world model, described by the company as a general-purpose physical intelligence.
- CaliBench 2026-08-10 released (benchmark)Odyssey benchmark that tests whether world models are physically calibrated.
- Agora-1 (multi-agent world model) 2026-05-18 research preview (public demo)Odyssey world model in which several people or AI agents share and act in the same generated world at once.
- Starchild-1 2026-05-18 announcedOdyssey world model the company describes as the first real-time multimodal world model.
- PROWL-1 and PROWL-2 2026-05-12 announced (research)Odyssey research models that learn to act inside simulated worlds; the second learns simulation and decision-making together.
- Odyssey-2 Max 2026-04-21 private beta (partners)Larger Odyssey world model aimed at more physically accurate simulation for robotics, games and defence partners.
- Odyssey-2 Pro 2026-01-23 released (API with SDKs)Upgraded Odyssey world model offered to developers by API for interactive video streams.
- Odyssey-2 2025-10-27 released (browser demo; API to follow)Odyssey's general-purpose world model that generates interactive AI video instantly.
- Odyssey-1 2025-05-28 research preview (public demo)Odyssey's first real-time "playable" world model: AI video that a user can steer as it is generated.
- Explorer (Odyssey image-to-world model) 2024-12-18 research (early; production trial)Odyssey's first model, which turns a single image into a detailed 3D world.
Key facts and news (3)- Founders
- Oliver Cameron — co-founder; Jeff Hawke — co-founder; board member Ed Catmull (Pixar co-founder) one source eweek.com/news/odyssey-ai-interactive-video/
- Models released
- Explorer (18 Dec 2024), Odyssey-1 (28 May 2025), Odyssey-2 (27 Oct 2025), Odyssey-2 Pro (23 Jan 2026), Odyssey-2 Max (21 Apr 2026), PROWL-1 (12 May 2026), Agora-1 and Starchild-1 (18 May 2026), CaliBench (10 Aug 2026), Odyssey-3 (15 Sep 2026), Agora-2 (24 Sep 2026), PROWL-2 (1 Oct 2026) one source odyssey.systems/writing
News - 12 February 2026 "Odyssey Announces Investment from NVentures & Samsung Next" — https://odyssey.systems/investment-from-nvidia-and-samsung (post not opened; not in deals_known — lead to verify as a funding deal) one source odyssey.systems/writing
Labs Round 8 (31 Oct 2026), from the researchers' rows; one-source items are marked. | | gov.uk/government/news/ai-firms-pioneering-drug- · press.airstreet.com/p/odyssey-series-b · techcrunch.com/2026/06/17/world-model-maker-odys · the-decoder.com/amazon-nvidia-and-amd-bet-310-mi | one source |
| PRAGMA (Revolut banking foundation model)Chips, models and products | Revolut's foundation model for sequences of banking events, trained on anonymised customer histories. scaling in real time (per NVIDIA)Key facts Pre-trained on 26 million anonymised user histories (CIC repost of NVIDIA, 13 Oct 2026); NVIDIA says it moved from research to real-time scaling | | blogs.nvidia.com/blog/uk-ecosystem-deployment/ · cic.vc/nvidia-announces-new-investments-in-the-u | one source |
| Safe IntelligenceCompanies · also on Safety, Assurance | Sells software that checks how a trained AI model behaves across very many possible inputs, to find hidden weak points before it is used.Profile- Founded
- Not stated (spin-out from Imperial College London)
- Founders
- Alessio Lomuscio — founder and CTO (Professor of Safe Artificial Intelligence, Imperial College London)
- Chief executive
- Steven Willmott — from Feb 2025 (earlier led product and engineering; founded 3scale, acquired by Red Hat)
- Chief scientist or research lead
- Alessio Lomuscio — CTO
- Hq
- London (England: London) per release dateline; other_offices: not stated
- What it does
- Safe Intelligence sells tools that check whether AI systems behave as intended. It began with formal verification and "robustification" of machine-learning models (vision and tabular), and in April 2026 launched Spec27, which writes tests from a specification and validates AI applications and agents before and after deployment.
- Research focus
- Formal verification and robustness of neural networks and tree ensembles; specification-driven validation of AI agents.
- Models and products
- Spec27 — "automated specification-driven validation platform for AI applications and agents", early access — 8 Apr 2026; verification and robustification platform and command-line tools — Feb 2025
- Customers or partners named
- Not stated (target sectors: finance, mobility, robotics, aviation)
- Selected papers
- Scalable Neural Network Geometric Robustness Validation via Hölder Optimisation (Y. Zhang, P. Kouvaros, A. Lomuscio) — NeurIPS 2025 — Dec 2025 — https://openreview.net/pdf?id=Va1IHqY3jG
- Rounds found
- 20 Feb 2025 — seed — £4.15m — Amadeus Capital Partners — OTB Ventures, Vsquared Ventures (already DEAL-LAB-2017 / DEAL-SAFE-002)
- Status now
- Trading; launched Spec27 in early access in April 2026.
Labs Round 8 deep profile (31 Oct 2026), as the researchers wrote it; confidence: one source. Sources - safeintelligence.ai/safe-intelligence-raises-4m-
- safeintelligence.ai/introducing-spec27-spec-driv
- safeintelligence.ai/scalable-neural-network-geom
| London (registered office N8 0RL) | Companies House · find-and-update.company-information.service.gov. · find-and-update.company-information.service.gov. · uktech.news/ai/safe-intelligence-raises-4m-20250 · safeintelligence.ai/safe-intelligence-raises-4m- | verified |
| Starling Bank banking foundation modelChips, models and products | Starling Bank is building a banking foundation model trained on Civo's UK sovereign cloud. in developmentKey facts No name, size or date published on the pages read | | blogs.nvidia.com/blog/uk-ecosystem-deployment/ · cic.vc/nvidia-announces-new-investments-in-the-u | one source |