A business owner told me recently that he wanted to "switch to the open source model" because his team had heard it was free and just as good as what they were paying for. Neither part of that sentence was quite right, and untangling it is worth doing, because the open-weight versus closed-model question is one of the more consequential decisions a small business can make about its AI setup, and almost nobody explains it in plain terms.
Here's the actual distinction. A closed model — Claude, GPT, Gemini, and similar — lives on someone else's servers. You send a request over the internet, their infrastructure runs it, and you get an answer back. You never touch the underlying model file, and you never could even if you wanted to; it's locked inside a company's data center. An open-weight model — Llama, Mistral, DeepSeek, Qwen are the names you'll hear most — is different in one specific way: the company that trained it publishes the actual weights, the enormous file of numbers that constitutes the trained model, and lets anyone download it. You can put that file on your own hardware, or rent hardware from a cloud provider, and run the model yourself, entirely outside that company's infrastructure.
That's it. That's the whole distinction. It sounds small, but it's the difference between renting an apartment and owning a building, and it matters a lot in practice. What it does not mean, despite the name "open source" getting thrown around loosely, is that the training data is public, that the training code is public, or that you could reproduce the model from scratch if you wanted to. In almost every case, the company still keeps the recipe — what data went in, how it was filtered, what the training process actually looked like — as a trade secret. Open-weight means "here's the finished cake, take it home." It does not mean "here's the recipe." Calling that "open source" the way we'd use the term for, say, an open-source database, is a bit of a stretch, and it's worth knowing that going in so you're not surprised later.
So why would a small business care about any of this? A few real reasons, and they're worth taking seriously rather than dismissing as engineer trivia. The biggest one is control over data. If you're running a model on your own servers, your customer data, your contracts, your internal documents never leave your building or your private cloud account — nothing gets sent to a third party's API at all. For a business in a regulated industry, or one handling data it's genuinely nervous about, that's not a nice-to-have, it's often the whole ballgame. A second reason is customization: you can take an open-weight model and further train it on your own material in a way that's much harder to do with a closed model behind an API. A third is insulation from a vendor's decisions — if a closed provider changes pricing, deprecates a model, or throttles access, you're at their mercy. Own the weights and nobody can take the model away from you or change the deal on you mid-stream.
Now the honest part, the part that gets glossed over in every excited LinkedIn post about "ditching the API and going open source to save money." Downloading a model file is free. Running it is not. These models need serious hardware — the kind of GPUs that cost real money to buy or rent, and the bigger, more capable open-weight models need a lot of them running continuously. Then you need someone who actually knows how to deploy the thing, keep it running reliably, patch it, monitor it, and troubleshoot it when it breaks at 11pm on a Friday, which it will. That's not a task you hand to whoever's good with Excel. That's a specialized skill, and it either means hiring for it, which is expensive, or contracting it out, which is also expensive, or you personally becoming reasonably fluent in infrastructure you didn't ask to learn.
This is the myth I most want to put a stake through: "open source means free." It does not, not for a business trying to run this in production. The model weights cost nothing. Everything around them — the servers, the electricity, the redundancy so it doesn't fall over, the security hardening so it isn't an open door into your network, the person or team keeping it alive — costs real money, often more per month than a small business would ever spend calling a closed API for the same workload. I've seen a client nearly commit to a self-hosted setup because the model itself was free, without anyone pricing out the GPU bill or the fact that nobody on staff knew how to keep a production model server running. That's the trap. The free part is the smallest part of the actual cost.
So when does open-weight actually make sense for a business like the ones I work with, which is to say small, lean, without an engineering department? Honestly, less often than the hype suggests. It makes sense when you have a genuine, specific data privacy requirement that a closed API can't satisfy no matter what contract you sign — certain healthcare, legal, or financial situations land here. It makes sense when you have a very high, very steady volume of a narrow, repetitive task, high enough that the math on owning versus renting compute actually tips in your favor over time. And it makes sense when you need to fine-tune a model deeply on your own proprietary material in a way an API just won't let you do. Outside of those cases, for most small businesses doing most tasks, paying a closed provider by the token and letting them handle all the infrastructure is not a compromise — it's the more rational choice, and I say that as someone whose job would frankly be easier if everyone believed the opposite.
There's also a middle path worth knowing about, because it's the one most small businesses that do want open-weight benefits actually land on: hosted providers who run the open-weight models on their own infrastructure and offer them to you as an API, same as a closed model, just with an open-weight model underneath. You get some of the benefits — often better pricing, sometimes better data terms, no vendor lock to a single model family — without taking on the server-management burden yourself. It's a reasonable compromise, and it's usually a more realistic starting point than buying GPUs.
The bigger lesson here isn't really about which acronym to prefer. It's that "open" in AI is a spectrum, not a light switch, and every point on that spectrum comes with a real bill attached somewhere — in dollars, in engineering hours, or in both. Anyone selling you a version of this where you get total control and it's also free is selling you something that doesn't exist. Figure out what you're actually protecting or optimizing for first — data privacy, cost at scale, customization, independence from a vendor — and then work backward to whether open-weight earns its keep for your specific situation, because for most small businesses, in most cases, it doesn't, and that's fine.
I write about this kind of thing regularly at 013labs.com, mostly because I keep running into business owners who've been sold a version of "open source AI" that oversells the free part and undersells the engineering part, and somebody ought to say the boring, accurate version out loud.