Blog
Notes from running inference on our own iron.
Engineering voice, no marketing. What we think about jurisdiction, open weights and operating GPUs, and what we learn doing it.
Inference has a jurisdiction
A prompt is data, and data is processed somewhere, by someone, under some law. Why we think that question decides more than any benchmark, and how Zarx answers it.
Why we serve open-weight models on iron we own
Owning the servers and serving open weights are the two choices that make the sovereignty claim a fact instead of a clause. What each one costs, and why we made them anyway.
What OpenAI-compatible means in practice
Compatibility is a promise about your code, not a marketing label. Exactly which endpoints, events and behaviours Zarx implements, what it does not, and how to verify it in an afternoon.