You’ve probably noticed AI companies split into two camps — some guard their models like trade secrets, others publish the whole thing for free download. This isn’t a minor technical detail. It’s one of the defining arguments in AI right now, and it affects who gets to build on this technology at all.
What “open” actually means here
Remember the “model” from earlier — the recipe the AI taught itself, made of billions of learned numbers called weights. A closed model keeps those weights locked away; you only ever talk to it through a company’s app or API, like ordering from a restaurant menu without ever seeing the kitchen. An open-weight model hands you the actual recipe file — you can download it, run it on your own hardware, and tinker with it directly.
the case for each side
The case for open
- Anyone can inspect, audit, or improve the model — not just one company's staff
- Startups, researchers, and small teams get to build without paying rent to a gatekeeper
- Runs offline, on your own hardware, with full data privacy
The "public library" argument.
The case for closed
- Easier to prevent obvious misuse when the company controls exactly how the model can be accessed
- Funds the genuinely huge cost of training — someone has to pay for it
- The company can pull or patch it fast if something goes wrong
The "responsible gatekeeper" argument.
Neither side is simply right — it’s a real tradeoff between access and control, and reasonable people land in different places on it.
the honest catch with "open"
“Open” doesn’t always mean “open”
Here’s the part that trips people up: some models marketed as “open” only release the finished weights, not the training data or the exact process used to build them — so you can run the recipe, but you can’t fully verify what went into it. That gap between “open-weight” and “fully open-source” is a real point of ongoing debate in the field, not just marketing pedantry.
where you've met each type
- Closed, app-based — Chatting through a company’s official app or website
- Open-weight — Models you can download and run on your own machine
- Hosted open models — Companies that host and serve open-weight models for you