Open models have caught up with commercial ones on many tasks, and an H100-class GPU rents for a couple of dollars an hour. That makes it tempting to run your own model and stop paying per token. Sometimes this really is cheaper and more convenient. Sometimes it is a way to pay three times as much for the same thing and get a server to look after as a bonus.
This series is about telling the two apart and doing it properly. First we work out when your own model makes sense at all. Then we learn to calculate which GPU a model needs, look at which models and engines are worth running right now, compare RunPod and Vast.ai, stand up our own endpoint with an OpenAI-compatible API, run image and video generation separately, and finally count the money.
There is little code in the series, only the commands you actually need to run. The substance is in the decisions: what to choose, why, and what it will cost. Prices and versions are as of early November 2026, and new open models come out every week, so the articles carry a feed of related releases.