What's the difference between open-source and closed-source LLMs?

Question

Accepted Answer

Closed-source LLMs like GPT, Claude, and Gemini only let you access the model through the company's app or a paid API, while open-source LLMs like Llama and Mistral let you download the actual trained model and run it on your own hardware.

The "open-source" label is a little loose — what actually gets released is the model's weights (the trained parameters), not the training code or data behind it, so people increasingly say "open-weight" instead. Either way, having the weights means you can run the model locally, fine-tune it, or modify it however you want.

What you trade for what: closed-source models need no setup — you just call the API — but you pay per use, your data passes through the vendor's servers, and the vendor controls updates, pricing, and availability. Open-weight models need your own hardware or a hosting provider, but they're free to run after setup, your data can stay entirely on your own machines, and you control the version and how it's customized. Closed-source models usually lead at the frontier on raw capability; open-weight models are often close behind, sometimes a generation back.

Neither option is better in every case. Pick closed-source when you want the strongest results with no setup and don't mind paying per query or sending data to a third party. Pick open-weight when privacy, customization, or long-term cost control matter more to you than squeezing out the last bit of raw capability.

	Closed-source	Open-weight
Setup	None — just call the API	You need your own hardware or a hosting provider
Cost	Pay per use	Free to run after setup, but you cover the infrastructure
Privacy	Your data passes through the vendor's servers	Data can stay entirely on your own machines
Capability	Usually leads at the frontier	Often close behind, sometimes a generation back
Control	Vendor decides updates, pricing, availability	You control the version and how it's customized