Loading the next useful route.
AI research guide
Why Europe needs to compete in model development, with a closer look at 15 labs and a wider directory to explore.
Explore 15 AI labs with European bases. Find work to follow, models to test and questions to ask before joining a team.
For researchers, engineers, founders and operators deciding what to build, try or join next.
Ambition and useful work
Europe should help build the most capable AI systems. By frontier, we mean pushing beyond what the strongest current models can do, broadly or in a specific field. That means backing new architectures and specialist research while continuing to compete in broadly capable models. The expertise should carry into future teams, even when a particular research bet fails.
This takes compute, capital and experienced people, with uncertain returns. If you are building a product, use the technology that works for your customers. European models need to earn adoption on their merits. Learn from work in Europe, the US, China and elsewhere, then test, contribute or join a team doing the underlying research.
The 15 teams below are a starting selection, not a ranking of frontier leaders or an exhaustive list of European AI labs. They span released models, specialist research and two announced research programmes. Locations identify European bases, including the UK and Switzerland, not exclusive headquarters or ownership. A published model is not automatically open-source or cleared for your intended use: check its licence and access terms.
Judith Dada’s essays make the case for turning European optimism into action and point to openings in image, video and world models, robotics and new learning approaches. JMTE’s position is that we should pursue those openings while continuing to compete in broadly capable models. Specialist research is not a reason to give up that ambition.
Four teams to investigate
Start with the problem you want to work on. These teams are pursuing different routes to more capable systems, at different stages of evidence.
Paris
Language and reasoning models, including downloadable model files. Worth exploring if you want to build products around broadly useful AI.
Start here: Choose a currently supported Mistral model for your task, then try your examples in Studio or through the API. For research, the 2025 Magistral paper explains an earlier reasoning-model training approach; it is not the starting point for a new integration.
Paris · also operates outside Europe
An announced research programme on world models for reasoning, planning and action. For people interested in how AI might learn how the world behaves. The stated ambition is not a demonstrated result.
Start here: Read the team’s explanation of its approach, then browse its roles. Compare the questions it wants to answer with the work you want to do.
London
An announced research programme on learning continuously through experience. Relevant if reinforcement learning and new ways of acquiring knowledge interest you. Treat its claims about future capabilities as research goals.
Start here: Read David Silver’s founding note. Pick one claim you want to see tested as results emerge, and ask what evidence would change the research plan.
Paris and London
Models and agents that work through computer interfaces. A team to investigate if you care about software completing tasks for people.
Start here: Start with the hosted API quickstart, then test one complete workflow in an isolated test environment. Record completion, retries and human intervention. A desktop benchmark score is not a guarantee that your workflow will succeed.
Four teams to investigate
These are useful places to start if you want to try the work as well as read about it. Use material you have permission to upload, and check account or payment requirements before testing.
Freiburg
FLUX models for creating and editing images. Relevant to visual tools and creative workflows. The original browser playground has moved to a developer playground.
Start here: Use the current developer tools to try changing one element of an image while preserving the rest. Compare consistency across attempts, not just the best result.
Paris · nonprofit research lab
Open research in speech and multimodal AI. Its conversational speech model Moshi has research and code available to explore. Kyutai also describes Gradium as its commercial spin-off, offering a different route into voice-model work.
Start here: Talk to Moshi. Interrupt it, ask a follow-up and listen to how it responds. Then inspect the code or paper if the interaction interests you.
London and Warsaw · global company
Models for generating and recognising speech. Worth exploring if you work on audio, dialogue or voice products.
Start here: For generated audio, compare delivery and consistency on the same passage. For live conversations, use the model-selection documentation and test a complete turn, including interruptions and names or numbers your customers use.
Paris
Voice models focused on real-time speech generation and transcription. For people building spoken interactions where response time matters.
Start here: Try its speech demo in two supported languages and check response time. Builders can continue with the API documentation and test a full interaction, including interruptions.
Four teams to investigate
Measurements, tables and the physical world call for different tests from a chat interface. Look at what the model receives, what it predicts and how the team checks the result.
Zurich
OpenTSLM connects measurements over time with language, combining existing large language models with components that process signals. Relevant to sensors, wearables and other sources of time-series data.
Start here: Look at the signal, question and answer in the OpenTSLM examples and quickstart. Consider what you would ask about your own data. Research evaluations are not clinical validation.
Freiburg · part of SAP
TabPFN predicts from structured, tabular data. Relevant if you work with datasets organised into rows and columns. SAP completed its acquisition in July 2026 and says Prior Labs continues to operate as an independent entity within the group.
Start here: Choose hosted API access or a local research setup, checking the model’s permitted use. Compare predictions on data excluded from training with your current method.
Linz
Research into xLSTM architectures and time-series models. Worth investigating if forecasting, model memory or efficiency interests you.
Start here: Read the TiRex forecasting introduction. Follow the model documentation to try a notebook, and keep future observations out of the inputs when evaluating a forecast.
London
Models for autonomous driving and embodied AI. For people interested in how AI understands and acts in the physical world.
Start here: Watch the GAIA driving examples and see how generated scenes change. These are simulations; assess them separately from real driving performance.
Three teams to investigate
These teams connect model development with experimental science. A promising benchmark or generated molecule does not by itself establish clinical safety or effectiveness.
London and Lausanne
Predictive and generative models for drug discovery. A team to explore if you want to work across AI, chemistry and biology. Alphabet remains a backer alongside external investors.
Start here: Read the Drug Design Engine overview to understand the problems it addresses. The linked report goes deeper into how the models were tested on unfamiliar molecules and proteins.
Paris
Biological foundation models, including H-Optimus and M-Optimus. Relevant to pathology and work combining different kinds of biological data.
Start here: Read the H-Optimus-1 and STAMP case study. Follow the research question and reported results, then inspect how the evaluation separates development data from test data.
London · also operates in San Francisco
Generative models for molecular design, including Latent-X, and the Latent-Y drug-design agent. Research access to Latent-Y requires approval.
Start here: Watch the recorded Latent-Y campaign to follow a design example. Check which results were tested experimentally and what the research access terms permit.
Look past the announcement
Start with the work you want to become good at. Your route may be model research, infrastructure, integration, experimental science or operations: check the actual role requirements. A lab can be worth following while still being the wrong place for your next job.
Funding matters, especially for general-purpose frontier models. It can support more compute, training runs and experiments. But a funding total alone doesn’t tell you what a team can achieve. Data quality, research choices and how efficiently the team uses its compute matter too. Ask what compute and runway this team actually has, and what it plans to achieve with them.
A round, debt facility and future spending commitment are different things, and global financing does not tell you a European team’s budget.
For specialist openings, start with ELLIS Jobs. EURAXESS broadens the search to research jobs, funding and hosting opportunities across disciplines. Filter by field and country, then check each role’s requirements and location with the employer.
Your role
What would you work on in the first three months? How much is model research, infrastructure, integration or customer work? Does that match what you want to learn?
People
Who will manage you and review your work? What relevant research or systems have they delivered? Ask to meet your direct colleagues.
Research progress
What does the work add to its starting model or strongest relevant comparison? Check the test data, evaluation setup and compute budget. Ask which results others have reproduced.
Research practice
Who chooses research directions, and what can you publish? Ask how publication approval and authorship worked on a recent project. Discuss an experiment that failed and what the team changed next.
Compute and runway
What resources are available to this team now? What can it complete before needing more funding? How do researchers get access to experiments?
Useful work
Who uses the research or product? For an early programme, what milestone would show the approach is worth pursuing?
A base for the work you want
For this selection, Paris connects language models, world-model research, speech and biology. London connects reinforcement learning, agents, voice, driving and molecular design. Zurich connects Aionic’s time-series work with RAI’s robotics research. Freiburg and Linz are reminders to follow the actual lab even when it sits outside the cities you already know.
Before planning a move, check where the relevant team works, how often people are together and whether the role puts you close to the work you want to learn. A company can have a European office while the team you care about sits elsewhere. Compare the people, the work and the life you could build there, then pressure-test the practical setup.
Also consider European teams inside global labs. Google DeepMind, Microsoft Research Cambridge, NAVER LABS Europe and Cohere’s London research operation belong in a search for research work here.
From reading to doing
You do not need to investigate every team. Pick the next step that matches what you want from this guide. For hosted model tests, use material you have permission to share and check how the provider stores or uses it.
Join a lab
Shortlist three teams working on a problem you care about. Read one paper or release from each, inspect their roles and write down what you still need to ask. Make your application specific about the problem you could help solve.
Build a product
Choose one customer task and a small set of representative examples. Decide what success means for quality, response time, cost and failures. Compare a model with your current method on the same examples. Count retries and manual corrections. Then choose: keep your current approach, investigate a failure or run a small pilot.
Bring other expertise
Look for where your experience meets the research. Labs also need people who understand biological experiments, industrial systems, product integration and company operations. Start from actual role requirements.
Explore
Pick two teams and give each a proper afternoon. Read the research, try what is available and write down what remains unclear. Then look for a relevant talk or technical community nearby.
A shortlist of your own
Browse 154 AI labs, model companies, research centres and initiatives with European bases in the view-only directory. On a computer, create a temporary filter view in Google Sheets to explore locations, organisation types and research areas without changing anyone else’s view. Follow the research and careers links for the teams that interest you.
The guide profiles 15 teams; the wider directory is a curated starting point, not a census or a ranking of frontier labs. It includes the UK and Switzerland, European teams within global companies, and clearly marked historical or uncertain entries. Source-review dates apply to the cited evidence, not every field or current operating status. Check the team’s current work, location and openings before acting.
Keep exploring
For earlier research bets, explore the teams in SPRIND’s Next Frontier AI Challenge. For academic research, start with ELLIS’s research sites. Neither is a substitute for checking the specific team and work you want to join.
For the compute question, Epoch AI’s historical study of language-model pretraining attributed most of the gains it analysed to increased compute and training data, with uncertainty about the split from algorithmic improvements. It did not assess post-training improvements. Research choices matter too: DeepMind’s Chinchilla outperformed Gopher using the same training compute budget, with a smaller model trained on more data. These are useful studies to read, not a formula for choosing an employer.
Company links lead to the teams’ own research, products and hiring pages. They support the descriptions here; they are not independent validation of performance, an endorsement of an employer or a promise of an open role.
Stay close to the signal
Event picks, useful Europe updates, and the occasional interview or video.