If someone told you “just use an open model, it’s free,” read this first
It’s a pitch more business owners are hearing: skip the big AI vendors, download an open model, run it yourself, and stop paying by the token.
Sometimes that’s great advice. Open models, the kind whose files you can download and run on your own computers, have gotten very good. (If “open model” is a new phrase, our field guide AI Isn’t One Thing explains it in plain English.) Mozilla’s new State of Open Source AI report finds that “the best open model trails the closed leader by three points at 60% of the price.” And they’re popular: in August, eight of the ten most-used models on OpenRouter, a service developers use to send real work to AI models, were open.
Mozilla is an outspoken champion of open models, and says so. That’s what makes the report useful: it’s also candid about the fine print. “A case that hides its weak points is an advertisement,” as its CTO puts it.
So here’s the fine print, translated for the person who signs the checks. Four things to check before “free” becomes a line item.
Check 1: “Open” doesn’t mean “do whatever you want”
People use “open weights” and “open source” as if they’re the same thing. They’re not.
“Open weights” means you can download the model’s file. “Open source” traditionally means you get everything needed to understand and rebuild it, with broad freedom to use it. Mozilla looked at sixteen notable open model releases and found that none of them ships the full data recipe the official Open Source Initiative definition asks for. In Mozilla’s words: “’Open weights’ does not mean open source.”
What you get instead is a license, and licenses vary a lot:
- Some are refreshingly simple. Thinking Machines released its Inkling model under Apache 2.0, a well-known permissive license. Even there, Mozilla notes the model card attaches “a separate, changeable Acceptable Use Policy.”
- Some are custom. Moonshot’s Kimi K3 license, per Mozilla, requires a separate agreement if you offer the model as a service above $20M a year, and requires displaying “Kimi K3” in your product above 100 million monthly users or $20M a month.
Let’s be honest: most businesses reading this will never get near those thresholds. The point isn’t that the terms are scary. It’s that every open model comes with conditions, the conditions differ from model to model, and they can change between versions. Moonshot’s previous model used a different license, and Mozilla notes that “did not settle K3.” Someone should read the license for the specific model and version you’re using, the same way you’d read a software contract.
Check 2: Can you actually run it?
“Download it and run it yourself” skips a big question: run it on what?
The largest open models are enormous. Mozilla describes Kimi K3 as needing roughly 1.56 terabytes of model files and 64 or more specialized AI chips to serve — “open, but not runnable by most who hold it.”
Smaller models are far more practical. Some run on a single high-end data-center chip, and very small ones run on a laptop or a phone. The trade-off is capability: by Mozilla’s measure, the best model you can run on one server scores about ten points below the top of the field, and the best on a single graphics card about twenty-three points below.
For plenty of business tasks (sorting emails, pulling fields out of invoices, drafting routine replies) a smaller model is more than enough. For others it isn’t. The right question isn’t “is it open?” It’s “what’s the smallest model that does this job well, and what does it cost to keep it running?”
Check 3: Who’s going to keep it running?
When you rent AI from a vendor, they handle the servers, the updates, the uptime, and much of the security and compliance. Mozilla notes closed vendors typically come with things like SOC 2 and HIPAA “packaged by default,” plus “a counterparty to hold liable.”
When you run an open model yourself, that’s all yours. And it’s where projects actually struggle. In Mozilla’s developer survey, the top problems with open models weren’t about the AI being good enough; they were running costs, security and compliance, ongoing maintenance, deployment, and a lack of support. The model file is free. The ops team isn’t. (We dig into that in The AI Isn’t the Hard Part.)
Check 4: Where did it come from?
This one’s less about any particular model and more about not putting all your eggs in one basket.
Mozilla reports that of the eight open models in OpenRouter’s top ten in August, seven were built in China. Those models are strong and widely used, including by large Western companies. But Mozilla, for all its enthusiasm for open models, still flags the concentration risk: “An open commons supplied from one origin is a dependency.”
For a business, two practical questions come out of this. First, do your customers or regulators care where your AI comes from? In some industries, like healthcare, finance, government contracting, and defense-adjacent work, they may, and it’s better to know before you build. Second, since no major release ships its full training data recipe, you can’t fully inspect what went into any of them. Test the model on your own real tasks, and don’t take published scores on faith. Mozilla notes some launch benchmarks are self-reported by the vendor on its own setup.
So when does open make sense?
Plenty of times. It’s often the right call when:
- Sensitive data can’t leave your walls. Running the model yourself means it doesn’t have to.
- You have steady, high volume and per-token bills are adding up.
- Continuity matters more than convenience. Nobody can switch off a copy you hold (more on that in Rent Your AI. Keep Your Keys.).
- You need a specialist. Per Mozilla, PwC fine-tuned a small open model on the language of finance and runs it on its own hardware for hundreds of clients.
And it’s usually the wrong call when your volume is modest, you don’t have anyone to run it, and a rented model already does the job at a price you’re happy with.
What to do on Monday
If an open model is on the table, whether someone pitched it or you’re curious, run it through four questions before anyone buys hardware:
- What does the license require, for this exact model and version? Who read it?
- What does it take to run, and what does a smaller model that could do the job cost to run instead?
- Who keeps it running: updates, security, uptime, and the 2 a.m. phone call?
- Where does it come from, and does anyone you answer to care?
Good answers to all four mean open might be a smart move. If there are blanks, fill them before you commit, not after.
That’s the kind of question our Readiness review is built for: a written, plain-English verdict on whether a path is safe, real, and worth what it’ll take. Or start with a free half-hour with a senior engineer, and bring the pitch you were given.