Open, open-weight, and proprietary models

There are a few kinds of model, and the labels get muddled. Proprietary (or closed) models like Claude, GPT, and Gemini stay on the company's servers. You reach them through an API and pay per use, with no way to download or inspect them. Open-weight models like Llama, Qwen, DeepSeek, and Kimi hand you the finished model file to download and run on your own hardware, free of per-token cost, though you don't get the training recipe. Open-source goes further, giving you the code and enough about the data to rebuild and audit the whole thing.

The catch is that "open" gets used loosely. Most models people call open source are really open-weight: you can run and fine-tune them, but you can't fully reproduce them, and the license may still restrict how you use them. So when someone calls a model open, it's worth checking which kind, since it changes what you can do with it.

The distinction matters most when you're choosing what to build on. Reach for a proprietary model when you want the strongest general answer with no setup and don't mind the metered bill and sending data to someone else's servers. Reach for open-weight when you want to run it yourself, to keep data in-house or avoid per-token costs at scale, and you can handle the infrastructure.

If licensing or auditability is the real concern, that's where true open-source matters, since open-weight alone doesn't guarantee you can use it however you like. For most people most of the time, the practical question is simpler: a closed model through an API, or an open one you host.