A closed model like the ones behind most major AI products is only accessible via an API call to the provider's servers, the company never sees or controls the underlying weights. An open-source LLM releases those weights publicly, so a company can download the model and run it on its own infrastructure, or a cloud provider's, entirely under its own control.

This matters most for organizations with strict data residency or air-gapped deployment requirements, since data never has to leave their own infrastructure to reach the model. The tradeoff is that running and maintaining the model well, keeping it updated, optimized, and secure, becomes the company's own responsibility rather than the provider's.

Frequently Asked Questions

Why would a company choose an open-source LLM over a leading closed model?

Data residency or air-gap requirements that a hosted API can't meet, a need for deep customization of the model itself, or cost efficiency at very high volume are the most common reasons, not because open-source models are generally more capable.

Does open-source mean the model is less capable than closed alternatives?

Historically, the strongest frontier models have generally been closed, though strong open-source options exist for many tasks. The gap and its relevance to a given use case shifts constantly, so it's worth evaluating for the specific task rather than assuming open-source is always behind.

What's the real cost of running an open-source LLM yourself?

GPU infrastructure, the engineering time to deploy and optimize it, and ongoing maintenance to keep it updated and secure. For most companies this total cost exceeds the sticker-price savings versus a hosted API, unless volume is genuinely very high or a hard requirement forces the choice.

Can an AI agent platform support both open-source and closed models?

Yes, and this flexibility is increasingly common, letting a company route different tasks to whichever model, open or closed, best fits that task's requirements for cost, capability, and data handling, rather than committing entirely to one path.