A lab in Kigali and a start-up in Cape Town are choosing an open model to run themselves. Every one, from GPT-2 to DeepSeek-V3, is the same shape with six knobs turned, and a handful of numbers in its config file give the parameters, the memory cache and the GPUs it needs. Then forget the untied head, give every head its own keys and values, or size the GPUs by active parameters. This lesson needs JavaScript.