§1What everyone is saying this week
The loudest claim on the timeline this week: GLM-5.2 from Z.ai is as good as the frontier, it is open weights, so you run it locally and you become sovereign, and no lab or government can take it away from you. The model is real and genuinely strong, one of the most capable open-weight models anyone has shipped. The "run it locally and become sovereign" part is where the maths quietly leaves the room.
We are not picking on anyone. The enthusiasm is fair, open weights catching up to the frontier is a real story. But three different ideas got collapsed into one sentence, and a builder reading the hype deserves to know which bits are true and which bit is doing the heavy lifting.
§2The number the hype skips
GLM-5.2 is a roughly 744 billion parameter model. At FP8 that is about 744GB of weights just to load, before it has produced a single token, and closer to one and a half terabytes at full precision. No laptop holds that. No gaming PC holds that. No single GPU holds that. You are looking at eight or more H100-class cards, a data-centre rig that runs comfortably past two hundred thousand pounds to own, or real money per hour to rent.
"Run it locally" is true in the way "anyone can own a yacht" is true. The weights being free does not make the hardware free, and on a model this size the hardware is the whole cost. Free weights and free to run were never the same sentence.
§3Local does not mean sovereign for almost anyone
So when the thread says run it locally and become sovereign, the honest translation is narrower: if you already own a GPU farm, you can be sovereign. For everyone else, the realistic way to use GLM-5.2 is the exact thing the pitch was meant to free you from, calling someone else's API on someone else's server.
We tried it ourselves this week, so this is not armchair maths. Our own gateway refused to route it on data policy, and even setting that aside, we own no machine that could load the full model. Open weights gave us the right to run it. They did not hand us a box that can.
§4What open weights actually buy you
None of this makes GLM-5.2 overhyped as a model. Open weights are a genuine shift, just not the one the thread is selling, and what they buy you is worth having. Three things. One, price: over an API it lands around Haiku money, a fraction of frontier cost, which is a real lever for high-volume work. Two, optionality: you are not married to one vendor's roadmap or one vendor's outage, the weights exist and someone will always serve them. Three, the real sovereignty option, for the few who genuinely need it: anyone with a hard data-residency requirement can, if they fund the hardware, run it on a box they control and keep the data in the building.
That third one is real, and it matters. It is also a procurement decision with a capital cost bolted to it, not a thing you unlock by downloading a file. Sovereignty by open weights is an option you can buy, not a default you get for free.
§5The honest builder take
The reframe is simple. Open weights are a sovereignty option, not a sovereignty default. The hype collapses three separate claims into one word: you can run it (true), you can afford to run it (only at data-centre scale), and you are therefore sovereign (only if you paid for the rig). Pull those apart and the panic to "get sovereign this week" deflates into a sensible decision with a price tag.
For the average builder, the practical reality is that GLM-5.2 is a cheap, strong model you will call over an API, and that is genuinely good news. If you truly need the data never to leave hardware you control, that road is open too, and it costs what a GPU rig costs. Knowing the difference is the whole point. The model did not repeal arithmetic, and nobody became sovereign by downloading a file.