Open-weight AI models: what firms should know

Firm leaders rarely need to care which model powers a tool. There is one exception, and it matters: whether the model can run somewhere you control. That single property determines where client data travels, whether a vendor can change terms underneath you, and what happens to your system when a provider retires a version. The vocabulary is confusing on purpose, so here it is plainly.

Three categories, not two

Closed models are reachable only through the provider's service. You send data out, you get an answer back, and the model itself never sits on your infrastructure. Most well-known assistants work this way.

Open-weight models publish the trained model so you can download and run it on hardware you control, though the training data usually stays private. This is the category most relevant to firms, and the one most often misdescribed as open source.

Truly open models publish weights, training data, and the training pipeline. Mostly of interest to researchers, but worth knowing the tier exists so you can tell when a vendor is stretching the word open.

The distinction firms actually feel: where the data goes

With a closed model, confidential material leaves your environment to be processed. That is not automatically wrong. Reputable providers offer contractual terms about retention and training, and for many firms that is acceptable once reviewed.

But it is a decision requiring an informed decision-maker, and it is the one most firms make by default rather than deliberately. Somebody enables a feature, and matter content begins flowing to a third party under terms nobody read. The question is not whether closed models are safe. It is whether your firm chose.

What running your own model actually costs

Open-weight models are frequently oversold to cautious firms as the obviously safer path. The honest accounting is less flattering. You take on hardware or hosting, updates, monitoring, and a security posture for the system itself. You inherit the work a provider was doing for you.

For most professional firms under a hundred people, that overhead is not justified by the risk reduction, because a reviewed vendor agreement addresses the same concern at a fraction of the cost. The real cases for self-hosting are narrower: a contractual obligation that data never leaves your infrastructure, a regulator expecting it, or a client whose agreement requires it.

The question that is not about safety at all

There is a second reason to care, and it has nothing to do with confidentiality: continuity. A closed model can be deprecated, repriced, or altered in ways that change your system's behavior on the provider's timeline rather than yours.

If you have built a workflow your firm depends on daily, that is a real dependency. An open-weight model you host can be frozen at a version that works, and upgraded when you choose. For some firms that stability is worth more than the privacy argument that usually gets top billing.

What to ask, whichever way you go

The category matters less than the answers to five questions. Ask them of any AI vendor, and ask them of yourself if you are hosting:

• Where is our data processed, and where is it stored afterward?
• Is our data used to train or improve the model? Get this in writing.
• How long is it retained, and can we require deletion?
• What notice do we get before the model or terms change?
• If this provider disappeared in ninety days, what happens to the workflow we built on it?

Where the middle ground sits

Most firms land between the poles, and sensibly so. General work runs on a reviewed closed model under acceptable terms. The narrow set of genuinely sensitive processing, if any exists, runs somewhere more controlled. That split is usually cheaper and calmer than an all-or-nothing decision in either direction.

Getting there requires knowing which of your AI uses touch what. Firms that have never inventoried their AI usage cannot make the split, because they do not yet know what is flowing where.

How Aday Interactive, Inc. approaches the choice

We start from your obligations, not from a model preference. Client agreements, regulatory posture, and the actual sensitivity of what the AI touches decide the architecture. Sometimes that is a reviewed vendor relationship. Occasionally it is infrastructure you control. Often it is both, split by use case.

What we will not do is recommend self-hosting as a default because it sounds safer. It carries real operating cost, and a firm that takes it on without capacity to maintain it ends up less secure than one using a well-reviewed service. See our Fractional CAIO work for how the decision gets made.