Mozilla’s latest State of Open Source AI report argues that open-source and open-weight AI have moved beyond experimentation and become a significant part of the global AI ecosystem, driven by rapidly improving capabilities, lower costs, and greater control over model deployment.
The September 2026 assessment examines the state of open AI across capability, adoption, infrastructure, investment, regulation, and sovereignty, including where Europe is gaining ground — and where it continues to lag behind the US and China.
Open AI moves from promising to ready
In a letter accompanying the report, Mozilla CTO Raffi Krikorian offers several examples of open AI being developed in Europe. In Lausanne, researchers built an open medical model with the Red Cross, tuned to its humanitarian guidelines, with clinical trials being prepared at home 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.”
He says these aren't outliers. Open-source and open-weight AI have become one of the fastest-growing builder ecosystems in software history: on Hugging Face alone, 2.5 million public models, 13 million users, and a third of the Fortune 500.
On OpenRouter, where developers route real production traffic, open-weight models grew from a sliver to roughly a third of usage by late 2025. Six months on, the platform moves 25 trillion tokens a week — five times more — and the single biggest source of that traffic is an open model.
“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.”
Europe is in the open AI race, but China leads the frontier
The report finds that while Europe is present in open AI, it is not leading at the frontier. The strongest open models assessed are overwhelmingly Chinese, with Kimi K3, GLM-5.3, and Qwen 3.8 among the leaders. European player Mistral appears further down the capability spectrum with Mistral Medium 3.5.
Further, Mistral’s openness comes with commercial restrictions. Mistral Medium 3.5 uses a modified MIT licence with a revenue carve-out, illustrating the report’s broader warning that “open weights” and genuinely open-source AI are not necessarily the same thing. Open AI as a route to greater technological sovereignty.
The report positions open AI as a route to retaining local ownership and control of models, data, and infrastructure, rather than renting access from dominant providers.
One example is Switzerland, where a public consortium trained a national model using public supercomputers and released the weights, data, and training code, making the model available for others to copy and adapt. This is unusual.
Across the 16 notable releases Mozilla examined, none provided the complete data recipe required by the Open Source Initiative definition. Europe’s opportunity may lie above the model layer
Further, the bigger European challenge may be deployment rather than model capability. Globally, 79 per cent of developers surveyed 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 rather than raw model performance. This creates an opportunity above the model layer for European companies, particularly in deployment infrastructure, orchestration, compliance, security, and enterprise tooling rather than necessarily trying to build another frontier foundation model. Economics could also lower the barrier for European entrants.
Capable open models can increasingly run on relatively modest hardware, while Mozilla argues that more capability gains are coming from post-training rather than ever-larger pretraining runs. This could shift differentiation towards post-training data, reward design, and reinforcement-learning environments, where smaller companies can compete with less capital and compute.
The UK bets £500M on sovereign AI
The UK is backing the development of Lumen Sovereign, which developer Cosine describes as Britain’s first sovereign frontier AI model. The model is being trained entirely on Isambard-AI in Bristol, using compute awarded through the UK government’s £500 million Sovereign AI programme.
Thirteen organisations are involved in the design phase, spanning banking, defence, telecommunications, and professional services, including HSBC, Lloyds, NatWest, LSEG, BAE Systems, Babcock, BT, and the Alan Turing Institute.
However, the report notes that their involvement represents design-phase agreements rather than purchase commitments.
Lumen is being developed using proprietary datasets covering more than 30 regulated workflows and is intended for air-gapped deployment within customers’ infrastructure by the end of 2026.
Because customers must hold the weights themselves, such deployments exclude API-only providers — creating what Mozilla sees as a growing market for open-weight and on-premises AI.
The model is being developed using proprietary datasets covering more than 30 regulated workflows, to enable air-gapped deployment within customers’ own infrastructure by the end of 2026. This requirement means customers need to hold the model weights themselves, effectively excluding API-only providers from such deployments. The report also points to Cohere’s plans to roughly triple its UK office footprint as another indication of growing demand for sovereign AI in the country.
Is the EU AI Act keeping pace with open AI?
The report questions whether the EU AI Act's use of training compute as a proxy for systemic risk can keep pace with increasingly efficient models.
Sparse models can now achieve comparable capabilities using significantly less compute, potentially placing similarly capable models under different regulatory requirements.
Mozilla also notes that open source does not provide a blanket exemption from the AI Act. Companies that substantially modify or adapt open models for high-risk uses, such as hiring, credit scoring, or medical diagnostics, can themselves become responsible for meeting the relevant regulatory obligations.
Would you like to write the first comment?
Login to post comments