Why a European AI Startup eustella Is Betting on Chinese Models

At first glance it sounds like a contradiction, maybe even a PR problem: a European AI startup that sets out to build a sovereign agent for smartphones is, of all things, betting on models from China. Anyone who reads eustella that way hasn’t understood the logic of today’s AI market — and is confusing origin with control.

Just ten days ago we announced what we’re up to: with eustella, we’re building a European AI agent for the smartphone, designed to give more than 100 million European AI users an alternative to ChatGPT, Claude and the rest. You can find all the details here.

The reflex is understandable. For months, geopolitical headlines have dominated the debate: export controls, chip embargoes, security reviews, the struggle for digital sovereignty. In that climate, the decision to build open models from DeepSeek, Alibaba’s Qwen or Moonshot into a European product looks at first like a strategic blunder. On second look, it’s the exact opposite — it’s one of the few options that allow a European startup to compete in the AI race at all.

The real question isn’t where a model comes from. It’s who controls it.

A closed model from California, accessed through an API, leaves European territory with every single request — together with user data, contextual information and everything else that makes up a smartphone agent. An open model from China, whose weights are freely available, can be run on European infrastructure, audited, fine-tuned and, if necessary, taken offline entirely. Sovereignty doesn’t come from the national flag flying over the training cluster. It comes from having authority over the weights, the inference and the data.

This is the point the debate too often skips over: in the past twelve months, the Chinese labs have done what neither OpenAI nor Anthropic nor Google were willing to do — they’ve released high-quality models as open weights. DeepSeek-V3, Qwen, Kimi, GLM: a whole generation of systems that hold their own against the Western frontier in benchmarks and are still free to use. For a European startup, that isn’t an ideological question, it’s an infrastructural one. Either you take what’s openly available, or you build your products on the API of a US corporation that could double its prices tomorrow, switch the model off, or change its terms of service.

Why China is pushing so aggressively into open source

It’s obvious that Chinese companies aren’t opening their models out of altruism. It’s a strategic calculation, not a gift — and the motives are worth unpacking briefly:

A response to US export controls. Without access to the most powerful Nvidia chips, Chinese labs can barely keep up with OpenAI, Anthropic or Google in a pure compute arms race. Open source is the lever they’re using to build global relevance anyway — through reach and community rather than raw compute.

Ecosystem effects as an accelerator. Whoever opens their models gets feedback, fine-tunes, benchmarks and tooling from tens of thousands of developers around the world for free. For latecomers, it’s the fastest way to close the gap.

The business-model logic of the conglomerates. Alibaba makes its money from cloud and commerce, Tencent from gaming and WeChat, Baidu from search. The model itself doesn’t have to be monetized — it serves as a lead generator for cloud services or as a strategic asset. That’s a fundamentally different starting position from OpenAI or Anthropic, whose core business is the model.

Commoditization as a weapon. When Chinese labs hand out good models for free, they push down the prices and margins of GPT-5.4-class APIs around the world. It’s a classic move: you make the layer where your competitor earns money worthless. DeepSeek-V3 and R1 triggered exactly that effect — including the violent stock-market reaction against Nvidia and Western AI shares.

Soft power and standard-setting. Open models of Chinese origin running in products worldwide are also a geopolitical statement. They give China influence over technical standards, tokenizers and language coverage.

A talent signal in the domestic competition. China’s AI market is brutally contested — in the “war of a hundred models,” a top benchmark result on Hugging Face is also a recruiting and reputation tool.

Regulatory reasons. Consumer AI products deployed inside China have to pass a security review by the Cyberspace Administration. Open model weights used by developers around the world don’t fall under the same logic — open source is, in part, the regulatorily easier way to reach the outside world.

For eustella, these motives are secondary. What matters is the consequence: the interest of Chinese labs in commoditizing American closed-source models happens to coincide with the European interest in technological independence. You just have to seize the opportunity. By the way, we’re not betting exclusively on Chinese AI. Google, OpenAI and Mistral (and perhaps soon Meta) also offer good open-source LLMs that we can and will use — depending on whether they meet our users’ needs.

Chinese open-source models are, incidentally, quite good. Below is the current ranking from arena.ai of the best AI labs offering open-source AI:

And the security concerns?

They’re real, but they concern something different from what many people assume. Open model weights aren’t remote-controllable. They don’t phone home. They don’t contain backdoors that activate with the next update — an open model doesn’t get updates, it’s a file. What you do have to examine are bias, censorship behavior, gaps in training data, and how a model reacts to sensitive topics. These are solvable problems: through fine-tuning, through evaluation, through combining several models, through clear guardrails in the agent layer above. And it’s exactly there — in the orchestration and the product — where an AI startup creates its value anyway, not in the base model.

The real controversy, then, shouldn’t be that a European startup is using Chinese models. The real controversy should be that Europe has spent years talking about wanting to build its own foundation models, and ended up with a handful of labs that, despite billion-euro announcements, fail to deliver what’s long been available on Hugging Face. Mistral is the honorable exception, but one swallow doesn’t make a summer. As long as that’s the case, it would be reckless to reflexively reject the good alternatives to American API dependence.

We’re being pragmatic

That’s why eustella is doing the only pragmatic thing: the company takes the best open models the market has to offer — whether they come from China, the US or Europe — runs them under European control, and builds an agent on top whose value lies in the product, not in the base model. Anyone who finds that naive is confusing sovereignty with symbolic politics. And anyone who finds it risky should ask themselves what the bigger risk really is: an auditable open-weight model on European servers — or quiet dependence on three US corporations that together decide what AI is allowed to do in Europe tomorrow, and what it will cost.

The truly European answer to the AI race isn’t to choose between Silicon Valley and Beijing. It’s to take what’s open from both camps and build something of your own out of it. The same thing is happening in the energy sector: there, millions of people want to become independent of oil and gas and put their energy supply on a more sovereign footing. And in Europe, they’re doing it largely with the help of solar panels and batteries (home storage, cars) from China.

eustella is doing exactly that. That’s not a controversy. That’s our strategy.

If you want to be among the first 5,000 users of eustella, sign up for the beta phase via this link.