The Triumph of Open Weight AI – and How eustella Builds on It

Not long ago, one thing seemed settled: the world's best AI models were proprietary, locked behind the APIs of OpenAI, Anthropic, and Google – and anyone who wanted frontier intelligence paid frontier prices. In the summer of 2026, that certainty is history. Within a few weeks, two models from China have shattered it.

On June 16, Z.ai (formerly Zhipu AI) released GLM-5.2, a mixture-of-experts model with around 750 billion parameters under an MIT license. It scores 51 on the Artificial Analysis Intelligence Index – the best result any open-weight model has ever achieved, just a few points behind Claude Opus 4.8 (56) and GPT-5.5 (53). On GDPval, a benchmark built around real-world work tasks, GLM-5.2 even matched a GPT flagship for the first time. And it does all this at roughly one-sixth of the API cost.

Kimi K3 – and what comes next

Exactly one month later, Moonshot AI raised the stakes with Kimi K3: 2.8 trillion parameters, making it the largest open-weight model ever announced, with a context window of one million tokens. Independent rankings put K3 on par with Claude Opus 4.8 and GPT-5.5 – only the very latest closed flagships, Fable 5 and GPT 5.6, still sit ahead. In the Frontend Code Arena, where developers compare outputs in blind tests, K3 even claimed the top spot, beating every US model.

And then there's DeepSeek, the startup that already sent one shockwave through the Western world. It has raised billions in fresh capital to build its next open-weight model, expected to launch in the coming weeks or months. What it will deliver is anyone's guess.

The numbers tell a clear story. The gap between closed US frontier models and open models from China has shrunk from years to – depending on the estimate – a few months. What was exclusive yesterday can be downloaded today, used commercially, and run on your own hardware. For companies that want to break free of US cloud APIs, for researchers, for Europe's sovereignty ambitions, this is a historic opportunity. For the first time, frontier-level intelligence is a freely available good.

The flip side of openness

But an honest look at this development has to include the other half of the story. Open weights mean anyone can download them – and anyone can modify them. Safety mechanisms that providers train into their models can be stripped out through fine-tuning. A model that codes at Opus level will also write malware at Opus level if you let it. Phishing campaigns in flawless German, automated vulnerability hunting, social engineering at industrial scale – the tools for all of this no longer sit behind an API with abuse detection. They sit on computers around the world.

There's more: once weights are published, they cannot be recalled. No patch, no update, no kill switch. And models from China raise questions of their own – about training data, about built-in biases, about benchmark claims that independent auditors have so far only partially verified.

The answer, however, cannot be to demonize open weights or pretend they don't exist. The genie is out of the bottle, and it won't go back in. The real question is: who makes this technology usable – and on what terms?

Between open access and an open flank

This is exactly where European providers come in. What's needed is a layer between raw model weights and the people who use them: an infrastructure that captures the cost advantages and independence of open models without passing their risks along unfiltered. Hosted in Europe, under European law, with abuse detection, guardrails, and clear governance – at prices not pegged to the margins of US hyperscalers.

That is the gap where startups like eustella operate. Making open-weight models accessible in a sensible, affordable, and controlled way means the frontier intelligence that GLM-5.2 and Kimi K3 set free actually reaches businesses, public administrations, and private users – as a productivity tool, not a hacking kit for anyone who wants one. Openness and responsibility are not a contradiction. But they don't happen on their own. Someone has to build them.

The triumph of open models is real, and it is good for Europe – if Europe shapes it instead of just watching from the sidelines.