Mozilla's latest State of Open Source AI report reads as a verdict, not a forecast: open AI is ready for production. The numbers from CTO Raffi Krikorian's letter paint an ecosystem that has become infrastructure. Hugging Face lists 2.5 million public models and 13 million users, with a third of the Fortune 500. On OpenRouter, open-weight models grew from a sliver to roughly a third of production traffic by late 2025. Six months later, the platform is moving 25 trillion tokens a week — five times more — and the largest single source of traffic is an open model.

The benchmark comparison makes the point plain: in spring the best closed model scored 60 and the best open models 54, up from 22 a year earlier for the leading open model. The frontier still leads on the hardest problems, but for the workloads most builders actually ship, price, control and deployability decide.

The report finds Europe present but not leading. The strongest open models assessed are overwhelmingly Chinese: Kimi K3, GLM-5.3 and Qwen 3.8. European player Mistral appears further down the capability spectrum with Mistral Medium 3.5, and its openness comes with commercial restrictions — a modified MIT licence with a revenue carve-out. That distinction between open weights and genuinely open-source AI matters because it determines what a company can actually do with the model and how much control it keeps.

The report positions open AI as a route to local ownership and control of models, data and infrastructure rather than renting from dominant providers. Switzerland's public consortium trained a national model on public supercomputers and released weights, data and training code. Yet that is an exception: across the 16 notable releases Mozilla examined, none provided the complete data recipe required by the Open Source Initiative definition.

The most useful finding for Europe may be about production, not capability. Globally, 79 per cent of surveyed developers use open models, but only 51 per cent of open-model users get them into production, compared with 63 per cent for closed models. Mozilla attributes the gap largely to operational tooling and trust, not raw model performance. In other words, capable models exist; what is missing is orchestration, compliance, security and deployment infrastructure. This shifts the competitive opportunity above the model layer, toward companies that can build reliable deployment pipelines for regulated sectors. For teams evaluating on-premise deployment, AI-RADAR offers analytical frameworks on /llm-onpremise for navigating these trade-offs. Economics lowers the barrier further: capable open models can increasingly run on relatively modest hardware, and Mozilla argues that more capability gains are coming from post-training rather than ever-larger pretraining runs. Differentiation is moving to post-training data, reward design and reinforcement-learning environments, where smaller companies can compete with less capital and compute.

The UK offers the most concrete test. Its £500 million Sovereign AI programme is backing Lumen Sovereign, trained entirely on Isambard-AI in Bristol and described by developer Cosine as Britain's first sovereign frontier AI model. Thirteen organisations are involved in the design phase, including HSBC, Lloyds, NatWest, LSEG, BAE Systems, Babcock, BT and the Alan Turing Institute — though Mozilla notes these are design-phase agreements, not purchase commitments. Lumen is being developed with proprietary datasets covering more than 30 regulated workflows and is intended for air-gapped deployment within customers' own infrastructure by the end of 2026. Because customers must hold the weights themselves, API-only providers are effectively excluded from such deployments. That creates a growing market for open-weight and on-premises AI, where physical control of weights becomes part of compliance and security. Cohere's plan to roughly triple its UK office footprint is another signal of rising demand for sovereign AI.

The regulatory question is whether the EU AI Act's use of training compute as a proxy for systemic risk can keep pace with more efficient models. Sparse models can now achieve comparable capabilities using significantly less compute, potentially placing similarly capable models under different requirements. Open source also does not provide a blanket exemption: companies that substantially modify open models for high-risk uses such as hiring, credit scoring or medical diagnostics can become responsible for meeting the relevant obligations.

Europe's path may not be to chase the frontier, but to build deployment infrastructure it can control. That is less a story of a European model beating the Chinese leaders and more an ecosystem trying to turn open AI into a sovereignty asset, one self-hosted deployment at a time.