Open AI Models: What Really Changes for a Business

When a business looks into artificial intelligence, the question always comes back in the same shape: which model should we choose? And behind the question, a stubborn assumption: an open model must be the cheaper, lesser version — the one you take when the budget is tight, while you wait to afford the "real" model.

That assumption is out of date. In two years, the gap has closed. The best open models now compete in the same league as closed ones, a few points apart on the benchmarks, and sometimes ahead on specific tasks. That shift changes far more than a ranking: it moves the question from choosing a vendor to choosing an architecture.

For a business, that is not a technical detail. It is what decides whether your data stays with you, whether your tools keep working when a provider changes its prices, and whether you can plug AI into your own systems without asking anyone's permission.

An open model is no longer the cut-price option

For a long time, "open" meant "a step behind". You used it to prototype, for low-stakes tasks, or because the budget would not stretch. The closed model of the moment stayed the reference — the one you migrated to as soon as a project got serious.

The relationship has flipped. What was true two years ago is no longer true: the most recent open models reach a level of reasoning, writing and comprehension that is enough for the vast majority of professional uses. Drafting a quote, summarising a contract, sorting incoming requests, answering a customer in their own language, extracting data from an invoice: on these tasks, the difference between the best open model and the best closed one is often invisible in practice.

The remaining gap sits on the hardest tasks — long reasoning, complex code, chains of autonomous actions. But it narrows with every generation, and it narrows fast.

The point most people miss: the question is not "which one is strongest?" but "which one is strongest for what I actually do". A model that wins one point on an abstract benchmark brings nothing to a business whose real need is to sort emails and write accurate replies.

What "open" actually means — and what it does not

The word "open" covers very different realities, and this is where many businesses go wrong.

• Open weights: the model file can be downloaded and run on your own machines. You can install it, adjust it, use it without call limits. This is the most common meaning today.
• Open weights under a restrictive licence: the model can be downloaded, but its licence imposes conditions — company size limits, bans on certain uses, obligations to credit the publisher. Read it before deploying in production.
• Open source in the strict sense: training code, data and method published, usage rights without restriction. This is the rarest case, and often not the one people assume.
• The API of an open model: you use an open model, but hosted by a third party. Convenient, cheaper than a closed model — but your data leaves your premises, and you depend on a middleman again.

The distinction matters: a model with open weights under a restrictive licence does not offer the same freedom as a genuinely free one. And an open model called through an API does not offer the sovereignty you were expecting from it.

"Open" does not mean "free of charge", nor "free of constraints". It means: you can look inside, and you can run it yourself.

Three questions before deciding

Before choosing between open and closed, three questions are worth more than a performance comparison.

1. Where does my data live? If your data is sensitive — customer files, health data, financial information, intellectual property — the ability to keep everything on your own machines is not a comfort, it is a requirement.
2. What happens if my provider changes tomorrow? A closed model is a rental: the day the price doubles, a feature disappears, or the API is withdrawn, you have no recourse. An open model, even if it takes more work at the start, cannot be taken away from you.
3. Do I have the skills to run it? An open model does not install like an app. You need someone who can deploy it, update it and monitor its quality. Without that skill, an open model becomes technical debt.

The three answers do not always point the same way. A business with no sensitive data, no in-house technical skills and no need for fine tuning is often better off starting with a hosted service. A business that handles confidential data, or wants AI at the heart of how it runs, is often better off looking closely at open models.

What it changes in practice for a business

Talk of sovereignty stays abstract until it is translated into real situations. Here is what genuinely changes, on the ground.

Cost. A closed model is paid for by usage: every call has a price, and that price climbs with volume. An open model is paid for in hardware and engineering time: once installed, every extra request costs almost nothing. Below a certain volume, the hosted service stays cheaper. Above it, the open model pays for itself — sometimes very quickly. The line depends entirely on your usage profile, not on a general rule.

Customisation. An open model can be adjusted on your data: your trade vocabulary, your formats, your internal procedures. A closed model can only be customised at the margins, through the prompt or external layers. For a business whose trade has its own language — a workshop, a practice, a specialist shop — that difference shows up immediately in the quality of the answers.

Version control. Open model: you decide when to move to the next version, and you can roll back if it behaves worse. Closed model: the provider updates when it likes, and your tool can change behaviour overnight, without notice.

Continuity. If your closed provider shuts down, gets bought, or decides to refocus, your service stops on the date they choose. With an open model installed at your premises, the tool keeps running.

Data and compliance: the argument that weighs most

For many businesses, the decisive argument is neither cost nor performance: it is what you are allowed to do with your data.

When you send a document to a hosted AI service, that document leaves your business. Depending on the provider, the contract and the country where processing happens, it may be stored, analysed or used to improve the service. That is not necessarily a problem — but it is a decision that commits you, and one you must be able to justify.

An open model installed on your machines removes the question at the root: nothing leaves. For a practice handling personal files, for a business handling financial data, for any organisation under confidentiality obligations, that property is often worth more than any performance gain.

It is also a question of bargaining power. As long as your tools depend on a service you do not control, every new provider becomes another point of dependency. Eventually you are no longer choosing your tools: you are putting up with theirs.

What it means in the Basque Country

Our local economy is made of human-scale businesses, often family-run, often attached to their territory. That profile changes the picture.

On one hand, these businesses have no IT team available to install and maintain an open model. They need someone to handle it, turnkey, with a contact they can call.

On the other, their relationship with confidentiality is far more concrete than one might think. A customer file is a name, an address, a situation. An order book is the result of twenty years of work. These businesses do not want that information drifting around; they want tools that do the job without exposing what matters.

That is exactly where the open approach makes sense: a tool installed on site, adapted to the trade, keeping data inside and depending on nobody to keep running. Local does not mean small — it means: here, with us, under our control.

Should you start now?

Two symmetrical mistakes are waiting.

The first: waiting. "We will see when things settle down." The trouble is that nothing ever settles in this field, and waiting only lets competitors get ahead on tasks that are, in fact, already mature.

The second: switching everything at once, on principle. Installing an open model on every desk, with no identified use and no in-house skill: the surest way to conclude that "it does not work".

The right starting point is more modest and more useful: pick one repetitive, time-consuming, low-risk task and handle it thoroughly. Sorting requests, reviewing quotes, personalising a standard reply. You measure the time saved, check the quality, adjust — and decide what comes next with facts, not a hunch.

It is also the best test of whether open models suit you: on a real case, you see immediately whether the model performs well, whether sensitive data is an issue, and whether the team follows along.

How I actually work

I do not arrive with a principled preference for open or closed. I arrive with a method.

First, look at the real work: what takes time, what repeats, what blocks. Then, map the data: what can leave the business, what cannot. It is that map, not a model ranking, that decides the architecture. Finally, deploy a first case, measure it, and expand only what proves its worth.

Depending on the case, I recommend a hosted service, an open model installed on site, or a combination of the two: open for anything touching sensitive data, hosted for ordinary tasks. The right choice is almost never ideological; it is contextual.

If you want to see what this looks like on your own tasks, browse our articles or let's arrange a call. We look at your situation, and decide together.

In short

• An open model is no longer the poor relation: on most professional uses, it performs just as well.
• What matters is not the ranking, but four questions: where data lives, what dependency you accept, what customisation you want, what skills you have.
• For sensitive data, it is the decisive argument: installed at your premises, nothing leaves.
• In the Basque Country, the demand is not "the best model in the world", it is an adapted, discreet tool that keeps data inside.
• Start small: one task, measured, before expanding.