From sovereign infrastructure to sovereign capability: Europe in the open-weight AI race

Updated 20 July 2026. I use the term “open-weight” deliberately. Access to model weights is not the same as open-source software, and it does not necessarily provide transparency over training data, code, evaluation methods or governance.

Over the past few weeks, I have been struck by the pace at which Chinese AI laboratories have moved.

Z.ai has released GLM-5.2, a model designed for long-horizon coding and agentic work, with a one-million-token context window. Its weights are publicly available under the MIT licence. Moonshot AI has launched Kimi K3, a 2.8-trillion-parameter multimodal model aimed at long-running coding, knowledge-work and reasoning tasks; it is currently available as a hosted service, with full weights due to be released by 27 July. Alibaba has introduced Qwen3.8-Max-Preview, currently offered as a hosted preview for complex reasoning, coding and visual understanding. At the time of writing, its weights are not yet publicly downloadable. 

MiniMax has also released the weights of MiniMax M3, a natively multimodal model with approximately 428 billion total parameters, 23 billion activated parameters and a one-million-token context window. This follows the international impact of earlier DeepSeek and Qwen releases, which showed how quickly a downloadable model can attract fine-tunes, deployment tools, inference providers and downstream applications. 

These models are at different stages of release, and it would be wrong to describe all of them as open-weight today. Nor would I call any one of them simply “the most powerful”. Performance varies greatly between coding, reasoning, multimodality, long-context work, tool use, inference efficiency and cost. Parameter count is not a proxy for quality, and results published by model developers need independent verification.

Even with those qualifications, the direction is clear. Chinese laboratories are treating model availability—not only model performance—as a competitive instrument. The cumulative effect is to draw developers, infrastructure providers and application builders into their ecosystems.

This matters for Europe.

The strategic asset is not only the model

Discussion about frontier models often becomes a comparison of benchmark scores. That misses part of what is happening.

When a capable model is released with downloadable weights, developers adapt it, fine-tune it and integrate it into products. Cloud and chip providers optimise their infrastructure around it. Universities use it in teaching and research. Tooling, documentation and specialist expertise accumulate around it. Organisations begin to make long-term architecture decisions based on it.

Over time, the model becomes more than a technical artefact. It becomes the centre of an ecosystem.

The organisation—or region—behind a widely adopted model can influence:

• the tools and technical standards used by developers;
• demand for particular cloud and hardware services;
• the direction of downstream research and investment;
• the economics of inference and fine-tuning;
• the platforms on which future applications are built.

Open-weight models can reduce dependency at the application level because organisations are able to deploy them on their own infrastructure. But they can also create a broader ecosystem dependency when the models, tooling and technical roadmaps on which European organisations rely are principally developed elsewhere.

This leads to a more important question than whether Europeans can access advanced AI:

Will Europe mainly consume, deploy and regulate AI ecosystems created elsewhere, or will it retain the capability to build globally competitive ecosystems of its own?

The American picture is more mixed

The United States remains extraordinarily strong in frontier AI companies, capital, research infrastructure, advanced chips and global distribution. Its model-release strategy, however, is not uniform.

OpenAI has released gpt-oss-120b and gpt-oss-20b under Apache 2.0, providing downloadable reasoning models designed for tool use and local deployment. Meta’s Llama 4 family remains one of the most significant open-weight ecosystems, although Meta uses its own licence rather than a standard open-source software licence. Google’s Gemma 4 family focuses on multimodal and agentic deployment on local and edge devices. NVIDIA’s Nemotron 3 programme publishes not only weights but also training data and recipes for several models. 

At the same time, many of the most capable American commercial systems remain proprietary and are accessed through controlled services.

The emerging landscape is therefore more complex than “open China versus closed America”. Both countries contain a mixture of open-weight and proprietary approaches. Companies make different choices about where openness helps expand adoption, where control supports monetisation and where technical or safety considerations justify restricted access.

What they share is the ability to operate at scale: to train models, release them, distribute them globally and sustain the surrounding infrastructure.

Europe must develop that same capacity.

Europe has credible players—but not enough depth

Europe is not starting from zero.

Mistral AI has demonstrated that a European company can develop internationally relevant foundation models and compete for global developer adoption. Mistral Large 3 was released under Apache 2.0, and the company has continued to work across coding, speech, document intelligence, mathematical reasoning, physics and robotics. Its recent expansion beyond general-purpose language models is particularly important: European capability should not be defined only by chatbots. 

Other European actors are building valuable capabilities. Black Forest Labs has become internationally relevant in visual foundation models. Aleph Alpha is developing sovereign AI infrastructure and specialised enterprise systems. Tilde has used EuroHPC resources to train a 30-billion-parameter model designed around European languages. 

OpenEuroLLM is developing common European foundations for transparent multilingual models, including datasets, documentation, training code and evaluation infrastructure. The project has secured more than ten million GPU hours across EuroHPC systems and is working towards an 8-billion-parameter model, followed by a larger model using its strategic compute allocation. 

In June, the Commission selected the EUROPA consortium, led by the Italian company Domyn, as the winner of the Frontier AI Grand Challenge. The consortium is expected to develop an open-source frontier model covering all 24 official EU languages and will receive access to up to 2.5% of EuroHPC’s overall computing capacity for one year. 

These are real achievements. They should not be understated.

But one successful company, one consortium and a collection of research projects do not yet constitute a sufficiently deep European ecosystem. Europe has far fewer organisations capable of financing and sustaining successive generations of frontier-scale models than the United States or China.

We need more companies with Mistral’s level of ambition. We also need a broader range of European champions rather than expecting any single company to carry the continent’s strategic ambitions.

That means competitive teams working across language, coding, science, industrial engineering, multimodality, speech, healthcare, robotics and edge AI. It means models that are not only European, transparent and legally compliant, but also powerful, efficient, easy to deploy and attractive to developers outside Europe.

Trustworthiness matters. But trustworthiness without competitiveness will not produce sovereignty.

Sovereign infrastructure is necessary—but not sufficient

A recent contribution to this community proposed moving from permissive open-source practices towards Sovereign Open Infrastructure.

That post emphasised verifiable provenance, secure supply chains, institutional responsibility, public compute and protection against dependency on commercially controlled “open” technologies. It also raised an important concern: companies may use open releases strategically to capture developer ecosystems and create longer-term dependencies.

I agree with the direction of that argument. Open weights alone do not guarantee sovereignty.

A downloadable model may still have opaque training data, an uncertain maintenance future, restrictive licence conditions or dependencies on infrastructure and tooling controlled outside Europe. An organisation may be able to run the model today without having any meaningful influence over where it goes tomorrow.

However, I would add another dimension. Sovereign infrastructure must be connected to sovereign capability.

Europe should not build excellent, secure and well-regulated infrastructure whose principal role is to host models developed in the United States or China. We must also retain the knowledge, capital, compute and institutional capacity required to train, improve, distribute and maintain powerful models ourselves.

Sovereignty should not be confused with autarky. Europe will continue to use and benefit from technologies created elsewhere, just as European technologies should be used elsewhere. The objective is not isolation. It is credible choice: the ability to act without being structurally dependent on decisions made by a small number of external providers.

The EU has built important infrastructure. It must now turn it into an ecosystem.

The Commission has already taken meaningful steps.

Europe now has 19 operational AI Factories and 13 AI Factory Antennas, with AI Gigafactories under development. The Apply AI Strategy promotes technological sovereignty and a “buy European” approach, particularly for public-sector and open-source AI solutions. The new EU Open Source Strategy also recognises that Europe often contributes to open technologies while much of the resulting economic value is captured elsewhere. 

The challenge is no longer simply to announce more infrastructure. It is to connect that infrastructure with research, private investment, procurement, deployment and global distribution.

The AI Office’s recent Expert Forum report, based on contributions from more than 100 specialists, places the issue directly in terms of European competitiveness, sovereignty and security. Its central concern is not only whether Europe can access frontier models, but whether Europe can choose, control and benefit from them. 

I believe five areas deserve particular attention.

1. Make compute a pathway, not a prize

Frontier-model development is not a one-off training run. A serious team must be able to train a model, evaluate it, correct its weaknesses, conduct post-training, improve its efficiency and begin work on the next generation.

One year of compute can produce an important model. It does not necessarily produce an enduring model developer.

Promising European teams need milestone-based, multi-year pathways through the European compute infrastructure. A team that demonstrates technical progress at an AI Factory should be able to move towards larger resources without restarting an entirely new political and administrative process for each model generation.

Access should remain competitive and performance-based. But it must also be predictable enough for organisations to retain staff, plan research programmes and attract complementary private capital.

2. Back several teams—and give them room to scale

Europe should not select one official model and expect it to serve every language, industry and use case.

Nor should the Commission attempt to design model architectures centrally. Its most useful role would be that of strategic customer, co-investor, infrastructure provider and coordinator.

Europe needs a portfolio of competing companies, research organisations and open consortia. Support should be linked to measurable technical progress, efficient compute use, multilingual capability, genuine openness where appropriate and credible routes to deployment.

The portfolio should include both general-purpose models and systems designed around areas of European strength: engineering, advanced manufacturing, medicine, scientific research, energy, mobility and industrial robotics.

European companies must also be able to obtain the growth capital required to train several generations of models. Otherwise, public funding may pay for early technical development only for the resulting capability to relocate, be acquired prematurely or become dependent on a non-European platform at the scale-up stage.

3. Turn public demand into a first market

Compute alone does not create sustainable companies. Model developers also need demanding customers.

European public administrations, universities, healthcare systems and strategic industries could become early users of European AI when it meets clear requirements for performance, security, interoperability and cost.

This is where the “buy European” principle should become practical.

It should not mean purchasing an inferior system merely because of its origin. That would protect companies from competition rather than making them competitive. It should mean removing procurement structures that automatically favour incumbent global platforms, allowing credible European alternatives to be tested and giving public buyers the ability to consider resilience, portability and strategic dependency alongside immediate price.

Public procurement can provide revenue, operational feedback and reference customers. These are often as valuable to a growing technology company as another research grant.

4. Fund everything around the weights

A model checkpoint is not an ecosystem.

International adoption depends on much more:

• high-quality multilingual and sector-specific data;
• independent evaluation and safety infrastructure;
• post-training and fine-tuning capabilities;
• efficient inference, quantisation and model compression;
• developer libraries and integration tools;
• dependable European hosting;
• documentation, support and security maintenance;
• international distribution and community development.

Europe is often successful at funding research and less consistent at supporting the costly transition from a research result to a widely adopted platform.

This gap should be treated as an industrial-policy problem. Training a strong model that few developers can deploy is not enough. Europe must also make European models convenient to use, economical to operate and straightforward to integrate.

5. Be precise about “European”, “open” and “sovereign”

These terms are often used too loosely.

Open weights, open-source code, open training data and reproducible training represent different levels of openness. A model can provide downloadable weights while revealing little about how it was trained. Conversely, full publication may not always be appropriate for systems developed for sensitive applications.

“European” also requires more than geographical hosting. Relevant questions include:

• Who controls the organisation and its strategic decisions?
• Where is the intellectual property held?
• Who controls the licence and future releases?
• Which infrastructure dependencies could interrupt development or deployment?
• Does Europe retain the knowledge required to maintain and improve the system?
• Can users move the model and their applications between providers?
• Where is the economic value created and retained?

European funding and procurement programmes should publish clear criteria rather than relying on labels. Otherwise, “open-washing” and “European-washing” will weaken the purpose of the policies they are meant to support.

A European alternative—not European isolation

The recent pace of development around GLM-5.2, Kimi K3, Qwen3.8-Max-Preview and MiniMax M3 should be taken as a signal.

The lesson is not that Europe should copy China. It is not that every European model must be open-weight, or that Europe should turn away from American technology. Different models and applications will require different release and governance approaches.

The lesson is that model strategy has become industrial strategy.

Europe should aim to combine:

• globally competitive technical capability;
• genuine openness where it is safe and useful;
• transparent and dependable governance;
• European linguistic and cultural diversity;
• secure and interoperable infrastructure;
• strong companies capable of competing internationally;
• public-interest foundations on which smaller organisations can build.

Mistral and other European initiatives have shown that Europe can produce important AI technology. The task now is to make such successes less exceptional—and to ensure that European teams have the support required to become more numerous, more capable and more globally influential.

The objective should not be to win every benchmark. Nor should it be to reproduce the scale of every American or Chinese company.

It should be to ensure that Europe possesses sufficient independent capability to make real choices, protect strategic interests, serve its languages and industries, and contribute powerful alternatives to the global market.

I would be interested in the community’s views on three practical questions:

Should the EU support several competing frontier-model teams over multiple model generations, rather than concentrating its resources in one-off challenges?

How can European public procurement create a first market for European AI without lowering requirements for performance, security or value for money?

What minimum conditions should a model or provider meet before it is described as European, open or sovereign?

The next few years will help determine whether Europe is principally a market in which other regions’ AI ecosystems are deployed—or a place where globally important AI ecosystems are built.

Europe should be capable of the latter.