Here’s a fun fact: In China, AI knowledge is now so mainstream that food delivery companies like Meituan don’t just have the capacity, know-how, and incentive to train a proper mid-size LLM—they’re releasing it as open source. All weights included. No restrictions. Ready to run locally.

The new LongCat-Flash-Chat isn’t some toy project: 560 billion parameters total, but thanks to clever MoE architecture, only 18-31 billion are active at once. The thing beats GPT-4 and Claude in some benchmarks—especially impressive on agentic tasks.

I honestly can’t remember the last time I downloaded and tested an LLM from a European company (Mistral aside). And it’s not like the recipe for LLMs is some secret sauce.

The ingredients are all there

Software/Algorithms? All open source. There are countless tutorials, notebooks, and books showing you how to build an LLM from scratch.

Training data? Tons of it freely available, plus whatever data treasure companies have internally as their “edge.”

Computing power? Rent it in the cloud or run it in your data center.

Energy? Here’s where it gets tricky. Cheap electricity is the difference between “worth it” and “billion-dollar grave.” Germany? Not great on this front.

People with experience? Now we’re talking real problems. Good AI developers in the US now earn more than DAX board members over here. And you absolutely need people who’ve done this successfully before—folks who bring the success recipes, tuning tricks, and that parameter intuition you can’t learn from books.

Mindset? This is the biggest mismatch I see. Look, the German/European caution mentality has its place—technology impact assessment is fundamentally a good idea. But here’s the thing: if we don’t build up the know-how and experience extremely quickly, we won’t even be able to meaningfully weigh the opportunities and risks.

The litmus test

Every AI taskforce and expert council needs at least half its members to answer yes to these questions:

• How many AI agents have you developed?
• How many LLMs have you pretrained and fine-tuned?
• How many training datasets have you created?
• How many AI gyms and training environments have you designed?

That last one’s getting really exciting. Prime Intellect is building an open platform for RL environments with their Environments Hub—basically the foundation for the next generation of AI agents. Even Andrej Karpathy says environments are the new gold in the reinforcement learning era.

So where are we in Europe? When will we see the first open-source LLMs from German industry—not just from research? The tools are there. The knowledge is available. We just need the courage to actually do it.