Open-source AI narrows the gap as Europe pursues sovereignty infrastructure

Open source AI Europe efforts have a more capable model base to build on, according to Mozilla’s September assessment, which found the best open models approaching the top closed model on its cited benchmark. But the report’s data and Europe’s infrastructure plans point to a harder constraint: local control over a model does not solve dependence on foreign compute, cloud capacity, chips or power connections.

Mozilla uses “open AI” to cover open-source and open-weight models, not OpenAI the company. The distinction matters: releasing model weights permits users to run or adapt a model, while a fully open-source release also requires enough information, including the data recipe, to reproduce it. Tech.eu reported that none of the 16 notable releases Mozilla examined supplied the complete data recipe required under the Open Source Initiative definition.

How close are open models to closed AI models?

Mozilla’s assessment, as reported by Tech.eu, put the leading closed model at 60 and the best open models at 54 in spring 2026. The top open score had been 22 a year earlier. That comparison does not establish that open models match closed systems across workloads: Mozilla said closed models retained an edge on the hardest tasks. For many production decisions, operators should test representative workflows rather than choose on a leaderboard alone, as outlined in this guide to evaluating AI models for actual work.

The practical appeal is control. Companies can run models in their own environments, modify them and avoid relying solely on an API provider, a relevant option for regulated workloads. Yet adoption has not translated cleanly into deployment. Mozilla reported that 79% of surveyed developers use open models, while 51% of those users put them into production, compared with 63% for closed models. It attributed the gap mainly to operational tooling and trust, leaving room for European vendors in security, compliance, orchestration and enterprise deployment.

Europe is not leading the open-model frontier. Mozilla’s assessment identified Chinese models including Kimi K3, GLM-5.3 and Qwen 3.8 among the strongest it reviewed, while Mistral Medium 3.5 ranked lower. Tech.eu also reported that Mistral Medium 3.5 has a modified MIT licence with a revenue carve-out, another reminder that open weights and open source are different claims.

Europe’s bottleneck is physical infrastructure

Bruegel estimates that the EU has two gigawatts of AI compute capacity, about 5% of the global total, and says planned projects will barely change that share. Its policy brief argues that capital is not Europe’s main constraint; permitting speed and access to electricity grids are more consequential.

The EU’s AI Gigafactory model is designed to create demand as well as capacity. An Open Future analysis of the tender says operators must finance facilities up front, while the EU and participating states commit to buy compute access once the sites operate. The Commission estimates each facility needs at least €4 billion to €5 billion of investment. Public authorities are intended to serve as anchor customers, but the tender also allows global service providers and international industrial partners, including hyperscalers, to play that role. Their services may fall outside the scheme’s broader sovereignty conditions, a tension rather than a resolved policy outcome.

The UK provides a smaller national example. Cosine says its Lumen Sovereign project is Britain’s first sovereign frontier model, trained on Bristol’s Isambard-AI through compute awarded under the government’s £500 million Sovereign AI programme. It is intended for air-gapped deployment in customer infrastructure by the end of 2026. Thirteen organisations are participating in its design phase, though Mozilla notes those agreements are not purchase commitments.

Regulation adds another layer. The AI Act is taking effect in stages: general-purpose AI obligations already apply, while most high-risk rules were due from August 2026 or August 2027, according to a Commission proposal. Mozilla also questions the Act’s use of training compute as a proxy for systemic risk as models become more efficient. Open source is not a blanket exemption: organisations that substantially adapt a model for high-risk uses can take on the relevant obligations.

Open models can reduce dependency at the model and deployment layers. Europe’s sovereignty case will still depend on whether it can build and operate sufficient local compute, secure power access and create durable demand for that capacity.

This story draws on original reporting from Tech.eu.