A month ago, the United States suspended European access to a series of frontier models. The signal was clear: what you rent can be taken away. Many saw it as proof that Europe had definitively missed artificial intelligence. Today, one event qualifies that verdict.
What Z.ai just released
The Chinese lab Z.ai has released GLM-5.2: a 754-billion-parameter model, with 40 billion activated per token. Two things make it remarkable, its level and its license.
The level: on SWE-bench Pro, which measures the resolution of real coding tickets, it scores 62.1, ahead of GPT-5.5 (58.6). On FrontierSWE, 74.4%, neck and neck with Claude Opus 4.8 (75.1%), the best proprietary model. On tool use (MCP-Atlas), 77.0, just behind Opus 4.8 (77.8). An open model has joined the leading pack.
The license: weights published under MIT, documentation specifying "no regional restrictions". Cost around $4.40 per million output tokens, roughly six times less than its closed rivals.
Why this detail changes everything
The original risk was being cut off from a provider overnight. An open model answers that directly: you download the weights, run it on your infrastructure, the data never leaves your environment, and no one can revoke your access. What was frightening a month ago no longer has the same hold.
The European lab Pleias summarized the trap well: "you do not rent a substrate and call it sovereignty". That is right. But the usual conclusion, that we need a European champion or we are lost, misses the essential point. You can own the weights. And owning a near-frontier model now costs a fraction of what we imagined.
How to use it, concretely
Three paths, from the fastest to the most sovereign.
Test it in one line, via OpenRouter. The model is exposed under the identifier z-ai/glm-5.2. The API is OpenAI-compatible: you change the base URL and the model name, the rest of your code does not move. Pricing: $1.40 per million input tokens, $4.40 for output, one-million-token context window.
Go through the publisher, via the Z.ai API. Z.ai offers its own hosted access, also OpenAI-compatible.
Host it yourself, for sovereignty. The weights are public on Hugging Face (zai-org/GLM-5.2, FP8 version of about 750 GB). You serve them with vLLM or SGLang on your own servers (vllm serve "zai-org/GLM-5.2-FP8" --tensor-parallel-size 8). A GGUF version even makes it possible to run it on a large workstation. The data never leaves your infrastructure.
The Parrit vision: agnostic by design
That is the posture we deploy with our clients. We do not bet on one provider, we stay agnostic: the right model for each task, data that stays with the client, and the ability to switch the day a better one comes out. Last week, an American model. Today, an open Chinese model. Tomorrow, something else.
Our job is not to sell you a model, nor a slide. It is to deploy the thing that runs, on your infrastructure, with the freedom to change the substrate without breaking everything. The real question for an executive is not "who makes the best model", but "do you know what exists, and can you switch".