Learn · Beginner
Open vs. closed AI models — what "open weights" really means
When people argue about "open" versus "closed" AI, the crux is a single technical thing: the weights — the giant grid of numbers that is the trained model. A closed model keeps its weights secret; you can only use it by sending requests to the company's servers and getting answers back, like talking to a vending machine you'll never open. An open-weight model is one whose weights you can download, run on your own hardware, inspect, and build on. That distinction sounds dry, but it changes almost everything about who controls the technology and what you can do with it.
A spectrum, not a switch
"Open" gets used loosely, so it helps to be precise. Releasing the weights lets you run and adapt a model — that's what made LLaMA and then Llama 2 so pivotal: capable models that researchers and companies could finally run themselves, igniting a whole ecosystem of fine-tuned variants. But truly open science means more — the training data, the code, and the recipe, not just the final numbers. Projects like OLMo push for that fuller openness, releasing the ingredients so others can reproduce and study the model end to end, not just use it. And "open weights" is not the same as "open source" in the traditional software sense — many open-weight models ship under licenses with real restrictions. So the right question isn't "is it open?" but "how open, and under what license?"
An analogy
A closed model is a restaurant: the food is great, but you never enter the kitchen, you can't see the recipe, and you eat only what's on the menu, on their terms. An open-weight model is being handed the recipe and the ingredients: now you can cook it at home, tweak the seasoning, serve it to whomever you like, and learn how it actually works. The restaurant may be more convenient and polished — but the recipe gives you independence.
Why open weights matter
- Privacy and control. Running a model on your own machine means your data never leaves it — essential for sensitive work where you can't send everything to someone else's server.
- Research. You can't truly study a black box. Open weights let scientists probe how a model works inside — the foundation of fields like mechanistic interpretability.
- Cost and customization. You can adapt the model to your task with reward-based fine-tuning and run it without paying per request.
- Competition. Every serious open release is a check on the closed labs, keeping pressure on price and pace.
What makes it possible now
Open models used to lag far behind the best closed ones. They've caught up partly because of the scaling-law insight that a smaller, well-trained model can rival a much larger one — which makes a genuinely runnable open model competitive rather than a toy. The result is a steady stream of capable releases: a flagship open model with a huge context window, and even unconventional architectures arriving openly, like an openly-released diffusion language model that lets the whole community study a non-standard approach firsthand instead of taking a lab's word for it.
The honest tradeoffs
Open isn't strictly better — it's a different bargain. Closed models are often the most capable at the very frontier, come polished and maintained, and keep dangerous capabilities behind a gate. Open weights, once released, can't be recalled, and the same openness that empowers researchers also removes a safety lever. The debate is genuine and unsettled. But the trend is clear: more of the most interesting AI is becoming something you can hold in your hand rather than only rent — and that reshapes who gets to study, build with, and benefit from it.
The takeaway
When you hear a model is "open," ask the follow-ups: open weights, or open everything? Under what license? The answer tells you whether you're getting a recipe or just a fancier vending machine — and that, more than any benchmark, decides what the technology can do for you.