Call a model "open source" in a room of engineers and half of them will nod while the other half wince. The wince is correct. Most of the models the industry calls open source are not open source. They are open-weight, and the gap between those two words can decide whether you are allowed to deploy the thing at all.
What is the difference? An open-weight model is one whose trained weights are released publicly. You can download them, run them on your own hardware, and usually fine-tune them. What you do not necessarily get is the training data, the training code, or an unrestricted license. An open-source model, in the stricter sense the Open Source Initiative has defined for AI, releases the whole recipe: weights, code, data description, and a genuinely permissive license, so the system can be inspected, reproduced, and reused without conditions that claw back your freedom to use it.
Put simply: open-weight gives you the cake. Open-source gives you the cake, the recipe, and permission to sell your own.
Why the label is so often wrong
The confusion is not accidental. Releasing weights under a friendly-sounding license is good marketing, and open source carries more goodwill than open-weight. So a lot of widely used models get called open source when their licenses say otherwise.
Some model families ship weights under genuinely permissive terms. Others release weights under a custom community license with real restrictions: caps on usage at scale, acceptable-use policies, or limits on using outputs to train competing models. A smaller set of models are open-source in the full sense, publishing training details and data alongside the weights under a standard permissive license. All three get lumped together as open, and only reading the actual license tells you which one you are holding.
Open-weight gives you the model. Open-source gives you the model, its provenance, and the legal right to use it on your terms. Marketing blurs them. Your legal team should not.
Why this matters for on-prem and sovereign deployment
For running a model on your own infrastructure, here is the good news: open-weight is usually enough. If the weights are released and the license permits your use, you can deploy on-prem or air-gapped without needing the training data. Sovereignty of deployment does not require full open-source; it requires weights you are allowed to run.
The catch is the license, and for a regulated or sovereign buyer it is not a formality. A custom community license can restrict commercial use, cap usage, or impose an acceptable-use policy that a government or bank cannot accept. A model that is technically downloadable but legally encumbered is not a safe foundation for a system that has to survive an audit and a decade of operation.
Full open-source buys you two more things that matter at the high-assurance end. Auditability: when the training approach and data are documented, you can reason about provenance and what the model was exposed to, which is exactly the kind of question a regulator asks. And independence: a permissive, standard license means no vendor can change the terms under you later.
What to check before you commit
- Does the license permit your use, at your scale, in your jurisdiction?
- Are the weights enough for your deployment, or do you need training transparency for audit?
- Is the license standard and permissive, or a custom one that can change?
- Are you allowed to fine-tune, and to use the outputs the way you intend?
The practical takeaway
For most enterprise and government deployments, the right model is an open-weight one with a license that clearly permits your use, sized and quantized for your hardware. Full open-source is worth paying attention to when auditability and long-term independence are non-negotiable, which in sovereign and regulated work they increasingly are.
Either way, the decision is not made on the model's marketing page. It is made on its license and its benchmarks. The largest model is rarely the right one, and the most open model is whichever one you can legally run, inspect enough to trust, and keep running on your own terms.
FAQ
- Is open-weight the same as open-source?
- No. Open-weight means the trained weights are released and you can run them. Open-source, in the full sense, also releases the training code, data description, and a permissive license so the model can be reproduced and reused freely.
- Can I deploy an open-weight model on-premise?
- Usually yes, as long as the license permits your use. Weights are what you need to run a model on your own hardware. You do not need the training data to deploy.
- Why does model licensing matter for enterprises?
- Because a downloadable model is not the same as a usable one. Custom community licenses can restrict commercial use, cap scale, or impose acceptable-use terms a bank or government cannot accept.