Two decisions define Zarx. We serve only open-weight models, and we serve them on servers NPAW owns and operates. Neither is the path of least resistance. Both are the reason the rest of the story holds together.
Renting capacity moves the jurisdiction
The fast way to launch an inference service in Europe is to rent GPU instances from a cloud provider with a datacenter in Frankfurt or Madrid and put a model server on them. It is fast, and it changes less than it appears to.
The building is in Europe. The operator of the machines, the network and the control plane is whoever owns the cloud, and for the largest providers that is a company under US law with the legal exposure that comes with it. The contract will say the right things about data location. It cannot change who can be compelled, by whom.
NPAW has run its own data infrastructure since 2008, first for video analytics, in its own racks and on its own network. When we started running language models for our products, we bought the hardware, as we always had. The result is that the sovereignty claim on Zarx is not a paragraph in a contract. It is a set of servers, in Spain, operated by a Spanish company, on an IPv4 block registered to NPAW in the RIPE database. You can look it up.
What owning the iron costs
It costs money up front, and it costs a team. Datacenter-class accelerators are expensive, they need power and cooling that ordinary racks do not provide, and they need people who know how to keep an inference stack healthy at two in the morning.
It also means committing to hardware generations. Today Zarx runs on AMD Instinct MI300X accelerators, with 192 GB of HBM3 memory per GPU, serving models with SGLang and vLLM on ROCm. NVIDIA GB300 systems are arriving, and the Rubin generation is planned. Choosing hardware is choosing what you can serve well for the next few years, and there is no vendor to blame when the choice is wrong.
We accept these costs because they are the costs of being the operator, and being the operator is the point.
A closed model is a provider in the path
The second decision follows from the first. If the model can only be reached through its provider’s API, then the provider is in your data path, and no amount of owning servers changes that. Every prompt goes to them.
Open-weight models break that dependency. The weights are published; the license is public; the model runs wherever you can run the serving software. We can download a checkpoint, pin it, run it on our own GPUs and tell you exactly what version answers your requests. When the license allows, you can take the same weights and run them yourself. Nothing about Zarx is a lock-in you cannot walk out of.
Models we run today include GLM-5.2 from Z.ai under an MIT license, Kimi K2.6 from Moonshot AI under a Modified MIT license, and Gemma 4 from Google DeepMind under Apache 2.0. Family, context and license for each come from the public model cards, which we link on the models page. We do not guess at a context length when the card is not explicit; we leave it out.
Pinned checkpoints, honest catalog
Two operating rules come with serving open weights.
First, a model id maps to one published checkpoint. If we adopt new weights, they get a new id. Your evaluations stay valid until you decide to move; nothing changes under you.
Second, the catalog grows at the pace we can operate, not at the pace of announcements. A model reaches the catalog after it has run on our iron, with our serving stack, under load, for our own products. There will always be a louder release we have not adopted yet. We would rather be behind by a few weeks than serve something we cannot stand behind.
Open weights and sovereignty are the same argument
People sometimes treat “European” and “open” as two separate virtues. For inference they are one. A closed model served from a European datacenter still routes every prompt to its provider. An open-weight model served from a rented US cloud still puts a US operator between you and the machine. Only the combination, open weights on iron a European operator owns, removes the foreign provider from the path entirely.
That combination is Zarx. It is more work than the alternatives. It is also the only version of the promise we would be comfortable making to a bank.
If you want to run inference this way, request access and tell us which models you need.