The Future of Open AI Models

How Long Will What Is Free Today Remain Free?

America and China Lead, Europe Is Falling Behind

There are currently two centres of gravity in AI development: the American — represented by OpenAI, Anthropic, Google, and Meta — and the Chinese, led by Alibaba’s Qwen series, DeepSeek, and Zhipu’s GLM models. European presence in this competition is minimal: Mistral AI is the only significant European player, and its founder, Arthur Mensch, acknowledged in 2026 that ‘our models are not yet the best.’

This does not reflect a lack of European talent or ambition. The structural disadvantage is explained by limited access to chips, smaller capital-raising capacity, and a different regulatory environment. Despite export restrictions, Chinese players remain at the frontier: today’s Qwen3.6, DeepSeek-R1, and GLM-5.2 are all downloadable, locally runnable models whose benchmark performance approaches that of closed Western frontier models.

This means that local, offline AI systems can today reach the best performance globally available — not because European development is competitive, but because the leading open models are downloadable by anyone. Those who take advantage of this window are not waiting for future sovereignty; they are achieving it now, in the present.

The Trend: Frontier Models Are Increasingly Closing

Over the past two years, a clear direction has emerged across the entire AI market: the largest and most powerful models are appearing less and less often in open-weight — freely downloadable — form, and are increasingly available only as paid services.

In April 2026, Meta effectively abandoned the open development line of Llama and switched to the closed-source ‘Muse Spark’ model, available exclusively on Meta’s own platforms — Facebook, Instagram, WhatsApp, and Messenger. No downloadable weight file, no open licence, no local deployment option. This marks the end of an era in which one of the world’s largest technology companies regularly published its leading open models.

Alibaba released three closed models in succession in early 2026 — including the multimodal Qwen3.5-Omni and the agentic Qwen3.6-Plus — none of which can be downloaded. Mistral’s Large 3, the company’s 675-billion-parameter flagship, is also accessible only via API. Qwen3.7-Max and Plus — the most recent and most powerful members of the Qwen series at the time of writing — run exclusively via cloud endpoints, with no download option.

The pattern is consistent and unambiguous: leading developers are increasingly keeping their frontier models in-house, releasing only earlier, weaker generations to the public. Today’s best open model will be ‘the previous generation’ tomorrow — but at least it will remain available if it has already been downloaded and stored.

The Regulatory Risk: What Nobody Had Even Considered

Alongside business decisions, a new and previously little-discussed risk has emerged: the possibility of regulatory prohibition. In May 2026, a joint investigation by the Financial Times and the Alice AI safety research group found that a freely available tool called ‘Heretic’ could remove the safety guardrails from most of Meta’s and Google’s open models within minutes.

The immediate consequence: American, European, and British decision-makers began discussing whether open-weight AI models should be regulated as dual-use technology, similarly to certain defence equipment. This could potentially mean that the release of future open models would require authorisation, be restricted in certain countries, or be entirely prohibited.

It is important, however, to see clearly what this risk applies to and what it does not. Current export control legislation — including the US regulatory category ECCN 4E091 — explicitly treats already publicly released models as an exception: what has once been published is, in legal terms, outside the scope of export controls. A technical fact is even more important: once a model has spread widely — through mirror sites, archives, and personal machines — revoking its distribution is technically unfeasible.

For a fully offline system, this risk moves one step further toward complete immunity: there is no network connection through which a future prohibition could be enforced. The model that is on the machine stays there permanently — regardless of what happens in the regulatory environment.

Why the Open Market Will Not Close Entirely

Despite the trend, there is an important counterforce preventing open AI models from disappearing entirely — and this counterforce is not philanthropy, but geopolitical strategy.

For China, publishing leading models in open form is a national interest, not merely a business decision. By releasing competitive models in downloadable form, China is deliberately building global technological influence and standardisation positioning — particularly in countries that cannot afford expensive Western API subscriptions. The weekly token usage of Chinese models on OpenRouter surpassed that of American models in February 2026, and the gap has continued to grow since.

This means: as long as at least one major Chinese lab has an interest in open releases — because doing so captures market position from other, closed players — others also have an incentive to maintain competitive open models. Complete closure currently appears unlikely. But long-term reliance on Chinese models raises a different kind of trust and sovereignty question — and it is to this question that ArkeoAI’s approach provides an answer: the goal is not to always run the most recent model, but to keep what one’s work depends on on one’s own machine, under one’s own control.

The most probable scenario is the one already taking shape today: the market will become two-tiered. Frontier models will move behind closed, expensive paid services. Good but non-frontier models may remain in open form — though not necessarily under the best conditions, and not necessarily without limits.

Archiving as a Professional Decision

ArkeoAI’s own model archiving approach follows from this logic. Every freely usable, serious open model is available to us so that we can deploy the most appropriate one according to each client’s needs.

This is a concrete, technically verifiable risk-reduction strategy that simultaneously provides three guarantees. First: if a model becomes unavailable in the future — through a business decision, regulatory prohibition, or a developer closing off access — the archived version can be reloaded and put into operation at any time. Second: a client’s system will never be left without a model due to an update or outage, because the model is physically on the client’s machine or in ArkeoAI’s archive, not on a third party’s server. Third: archiving enables model comparison and on-demand substitution over time — tested on the same corpus, with the same questions, across different models and model-version combinations.

The Local Archive as a Strategic Tool for the Client

When ArkeoAI deploys a system at a client site, the archived models are not merely technical reserves — they are strategic decision options. If the client’s workflow changes (for example, new document types come into scope, or activities expand), a model change requires no additional licence fee, no cloud API connection, and no dependence on any foreign developer’s decision. The most suitable model can be retrieved from the archive, tested, and deployed with the client’s approval.

This is the flexibility that cloud-based solutions cannot offer: there, a ‘model change’ means the developer replaces the engine running in the background — without the client’s knowledge or consent. With a local archive, a model change is the client’s decision, because the client owns the model.

Predictability is one of ArkeoAI’s core commitments — and this commitment applies not only to a model’s behaviour, but to its availability. Because the model running on a client’s system today will still be there tomorrow — regardless of what is decided in a Silicon Valley boardroom, a Brussels committee, or a Beijing technology strategy session.