Open vs Closed Models

OpenClosed

Part of: Choosing the Right Model

Two teams build the same feature. One calls a closed model over an API and ships in a day. The other downloads an open-weight model and runs it on their own servers, slower to set up, but they own it. Neither is wrong. They made opposite bets on control, cost, and privacy, and the right bet depends entirely on the situation. What it is A closed (hosted) model runs on the provider's servers. You send a prompt over an API, you get an answer, you pay per token. You never touch the weights. An open-weight model has its weights published for download. You can run it on your own hardware (or a cloud machine you rent), change how it's served, and keep every byte of data in-house. The capability is yours to host. The split isn't "good vs bad." It's who runs the model and who bears the consequences. How it works The decision often comes down to total cost at your volume. Closed models charge per call; open models charge you fixed infrastructure plus a smaller per-call compute cost: There's a crossover point : below it, hosted is cheaper because you skip the fixed cost; above it, self-hosting amortizes that fixed cost across millions of calls. Why it matters Each side wins on different axes:

Challenge: Hosting Cost Crossover