According to Mozilla's latest State of Open Source AI report, open-source and open-weight models have transitioned from early-stage experimentation to become a meaningful force within the worldwide AI landscape. The shift stems from substantial improvements in performance, reduced operational expenses, and increased flexibility in how organisations deploy and manage these systems.

The September 2026 assessment evaluates open AI across six dimensions: capability, adoption, infrastructure, investment, regulation, and sovereignty. It specifically examines where Europe is making progress—and where it remains behind the United States and China.

Open AI moves from promising to ready

In remarks accompanying the report, Mozilla CTO Raffi Krikorian highlighted several instances of open AI development occurring within Europe. One case involved researchers in Lausanne collaborating with the Red Cross to develop an open medical model aligned with humanitarian principles, with clinical trials planned domestically and in Tanzania.

They built it themselves. They didn't ask permission, they didn't rent it — they own it, theirs to run, change, and keep.

Raffi Krikorian, Mozilla CTO

Krikorian stressed these are not isolated examples. Open-source and open-weight AI have emerged as one of the most rapidly expanding developer communities in software history: Hugging Face alone hosts 2.5 million public models, serves 13 million users, and counts roughly a third of Fortune 500 companies among its participants.

On OpenRouter, a platform routing actual production workloads, open-weight models expanded from negligible to roughly one-third of all traffic by the end of 2025. In the following six months, the platform processed 25 trillion tokens weekly—five times the earlier volume—with an open model representing the largest share of that traffic.

This spring, the best closed model scored 60 and the best open models 54. A year earlier, the leading open model scored 22. The frontier still leads on the hardest problems — and for the work most builders actually ship, where price, control, and deployability decide, the data does not say "promising." It says "ready." If you have been waiting for open AI to grow up, stop waiting.

Raffi Krikorian, Mozilla CTO

Europe is in the open AI race, but China leads the frontier

The report reveals that while Europe participates in open AI development, it does not occupy the leading position at the capability frontier. The most capable open models assessed come predominantly from China, with Kimi K3, GLM-5.3, and Qwen 3.8 ranking among the strongest performers. European contender Mistral ranks lower on the capability scale, with Mistral Medium 3.5 representing its offering.

Mistral's approach to openness carries commercial constraints. Mistral Medium 3.5 employs a modified MIT licence that includes a revenue carve-out, reflecting a broader observation in the report: "open weights" and truly open-source AI are distinct concepts. The distinction matters for organisations seeking technological sovereignty through open AI.

The report frames open AI as a mechanism for preserving local ownership and governance of models, data, and computational infrastructure, rather than depending on access leased from dominant technology firms.

Switzerland exemplifies this approach: a public consortium trained a national model using publicly available supercomputers and released the model weights, training data, and code, enabling others to replicate and modify it. Such comprehensive transparency remains uncommon.

Among 16 significant model releases Mozilla examined, none supplied the complete data and methodology required by the Open Source Initiative's definition. Europe's competitive advantage may therefore emerge not at the model layer itself.

The more pressing European challenge appears to be deployment rather than raw model performance. Globally, 79 per cent of surveyed developers employ open models, yet only 51 per cent of those developers successfully move open models into production settings, compared with 63 per cent for closed models.

Mozilla attributes this gap primarily to operational tooling and confidence in the systems rather than fundamental model quality. This disparity creates an opening for European organisations operating above the model layer—specifically in deployment infrastructure, system orchestration, compliance verification, security, and business-focused tooling. Rather than pursuing another frontier-class foundation model, European companies might find greater success in these supporting domains.

Economic factors could also lower entry barriers for European competitors. Increasingly capable open models can operate on relatively modest computing hardware, and Mozilla notes that additional performance gains are coming from post-training refinement rather than ever-larger initial training runs. This shift could move competitive advantage toward post-training datasets, reward mechanisms, and reinforcement-learning frameworks—areas where smaller organisations can participate with reduced capital and computational resources.

The UK bets £500M on sovereign AI

The United Kingdom is investing in Lumen Sovereign, which developer Cosine describes as the nation's first sovereign frontier AI model. Training takes place entirely on Isambard-AI in Bristol, drawing on compute resources allocated through the UK government's £500 million Sovereign AI programme.

Thirteen organisations participate in the design phase, representing banking, defence, telecommunications, and professional services sectors. These include HSBC, Lloyds, NatWest, LSEG, BAE Systems, Babcock, BT, and the Alan Turing Institute.

The report notes that these organisations have committed to design-phase participation rather than purchase guarantees.

Lumen development relies on proprietary datasets spanning more than 30 regulated workflows and targets air-gapped deployment within customers' own infrastructure by the end of 2026. Since customers must retain the model weights locally, this architecture excludes API-only service providers—creating what Mozilla identifies as an expanding market segment for on-premises and open-weight AI solutions.

Cohere's plan to roughly triple its UK office workforce signals additional momentum behind sovereign AI demand in the country.

Is the EU AI Act keeping pace with open AI?

The report raises concerns about whether the EU AI Act's reliance on training compute as a measure of systemic risk can adapt to increasingly efficient models. Sparse model architectures can now deliver comparable performance using substantially less compute, potentially subjecting similarly capable systems to different regulatory classifications.

Mozilla also clarifies that open source does not automatically exempt organisations from the AI Act. Companies that substantially modify or repurpose open models for high-risk applications—such as recruitment, credit assessment, or medical diagnosis—assume responsibility for satisfying the corresponding regulatory requirements.

Source: Tech.eu