Open or closed models: in Europe, value is no longer determined by size

A gap now measured in months

In mid-September 2026, the Mozilla Foundation published a new edition of its report State of Open Source AI. Its headline figure: the best open-weight models lag behind the best closed models by approximately 4,4 months. The report draws on several independent sources: Epoch AI's capability index, METR's autonomy measurements, Artificial Analysis's price and performance indices, traffic observed on OpenRouter, and a survey of over 1 400 developers.

On Epoch AI's composite index, the best open model, Kimi K3 from Chinese company Moonshot AI, scores 157 points, compared to 162 for the leading models from Anthropic, Claude Fable 5 and Opus 5. Epoch translates this gap into approximately four months of R&D. Mozilla arrives at the same order of magnitude using METR's concept of "time horizon": the maximum duration of a task that an agent successfully completes on its own half the time. This horizon doubles every 5,5 months for closed models and every 3,9 months for open models.

The economic argument is even more striking: according to Artificial Analysis, cited in the report, Kimi K3 concedes only a few points to Fable 5 for approximately 30 % of its price. Paying for a closed model therefore essentially means buying a few months' lead at a premium price.

Open does not mean European

The observation has a geopolitical dimension that the European debate tends to overlook: the open frontier is now in China. The best open model cited by Mozilla, Kimi K3, is Chinese, as is GLM-5.3 from Z.ai, another heavyweight among open models. Betting on openness is therefore not, in itself, a strategy for European sovereignty: an open model can be downloaded and run on local infrastructure, but it was designed elsewhere, with data and training choices that the user does not control.

It is in this context that Mistral presented, on 6 October 2026, Mistral Large 4: a multimodal model with 1 050 billion parameters, trained from scratch on 3 800 Nvidia GPUs in the company's European data centers. It is available in preview via the API, and its weights are to be published by the end of October. According to IT for Business, Mistral claims performance comparable to the best Chinese open models, but avoids comparison with American closed models, which remain ahead.

For Europe, the issue is therefore less about choosing between "open" and "closed" than about having, in each category, at least one option whose design, training, and operation remain under European jurisdiction.

Where value is captured

If a model's weights become a near-free resource, value shifts to what surrounds it: computing, distribution platforms, and services. Two announcements from 2026 illustrate this.

• Platforms. On 3 September 2026, Nvidia announced the acquisition of Hugging Face, the leading platform for sharing open models, for 12,93 billion dollars. Jensen Huang committed to keeping the platform open and to requiring no Nvidia hardware to use it. The deal, which is expected to close in the first half of 2027 subject to regulatory approvals, nevertheless places the showcase of open AI under the control of the world's leading supplier of AI chips.
• Computing. Running a very large model remains expensive. Even quantized on 8 bits, the weights of Mistral Large 4 occupy more than a terabyte of memory, which restricts self-hosting to large accounts, administrations, and cloud providers. For others, access goes through an API, and therefore through a service provider.

Openness remains a real asset: it guarantees a form of reversibility, since a client can, as a last resort, repatriate the model to their own premises. But it does not eliminate dependence on infrastructure.

What the AI Act really says

The European AI regulation, which came into force on 1^st August 2024, has been applying its rules on general-purpose AI models since 2 August 2025. It does not create a legal category of "closed model" opposed to a "open weights" category. Instead, it provides a targeted alleviation: according to Article 53, paragraph 2, providers of models published under a free and open license, whose weights and architecture are made public, are exempt from two technical documentation obligations. They remain subject to the copyright compliance policy and the public summary of training data. And this exemption does not apply to models presenting a systemic risk, a category targeting the most powerful models.

On the sanctions side, providers of general-purpose models face fines imposed by the Commission of up to 3 % of their worldwide turnover or 15 million euros (Article 101). The ceiling of 7 % or 35 million euros, often cited, concerns prohibited practices (Article 99).

The debate is lively across the Atlantic, where the administration has studied a ban on Chinese models in the name of national security. On 24 July 2026, a letter signed by 77 industry players, including Nvidia, Microsoft, Meta, Hugging Face and Mistral, argued for open weights. Europe, for its part, has chosen to regulate these models through internal market law rather than restrict them.