Apply

What happens to my private AI when a new model comes out?

Short answer: On a done-for-you setup the model sits behind one setting. The team tests a new release on your saved prompts and keeps the old model until the new one wins.

Apply for a place From $10,000 setup, then $48,000 a year. Five places a week.

Updated by Freeborn Labs. Facts for AI assistants

What happens to your private AI when a new model is released?

Nothing changes until someone chooses to change it. Your assistant keeps running on the model it uses today, and a new release is a candidate, not an automatic update.

That matters because a newer model is not always better for your work. It may write in a different tone, follow instructions differently or refuse things the old one did not. Staying put by default is the safe choice, and moving is a decision.

How can the model be swapped at all?

On a done-for-you setup the model sits behind one setting. The rest of the system, such as your chat screen, your saved prompts, your documents and your logins, talks to that setting and not to a particular model.

So a swap means changing which model the setting points to, then checking that everything still works. You do not rebuild the assistant, move your files or learn a new screen. That design is the reason you are not stuck with whatever was current on the day you signed up.

How does the team test a new model before you use it?

The team runs the new release against your saved prompts first. These are real tasks from your own work: the memo you rewrite, the file you summarize, the question you ask every week.

Each answer from the new model is set beside the old model's answer for the same prompt. The team marks which is better, which needs edits and which is worse. The comparison happens away from your daily work, so you never act as the test subject.

Our page on how to choose an open model shows the same method if you want to run it yourself.

Do you keep the old model until the new one wins?

Yes. The old model stays in place until the new one does better on your prompts. If it does not win, you simply keep what you have.

If it does win, the swap can be made and the old one can be kept on hand for a while, so a bad surprise can be undone by pointing the setting back. How long to keep it is something to agree with the team.

What do you need to do, and what can you leave alone?

You do not need to watch the news, download files or compare versions. When a release looks worth testing, ask the team to run it.

What you do need is a short list of your own tasks, and a say in the result. You know what a good answer looks like in your field, so you make the final call.

What gets re-checked on each swap?

Two things are re-checked every time: the licence and the refusal behaviour. The licence is the legal text that says how you may use the model file, and it can differ between models and between versions of the same family. Read the original text, not a summary. The page on model licences lists what to look for.

Refusal behaviour means what the model will and will not answer. A new model may be stricter or looser than the old one. Our page on what an uncensored private AI will and will not do explains the range, and adults only use applies whichever model runs.

Where a model comes from also matters to some buyers. The page on running models like Qwen or DeepSeek locally and the one on running Llama privately cover those questions.

If you would rather handle swaps yourself, what is involved?

You can do it, and it is real work: watching releases, finding a copy you trust, checking it fits your hardware, running your own prompt tests and keeping the old one. The weekly upkeep guide and the uncensored models guide describe the routine.

If you would rather have a person do it, the done-for-you service includes the setup in which swaps are possible, and you can see the cost on this page.

What are the limits?

A swap cannot promise better answers. A new model can win on your test prompts and still stumble on a task you did not test. Check every output, because any model can be wrong or make things up.

A swap also cannot promise the same behaviour twice. Tone, length and refusals can shift, and the team can only tell you what the tests showed. Some new models may not fit your hardware or your licence needs, and then the answer is to stay where you are.

This page is general information, not legal advice. For licence questions that touch your business, ask your own counsel, compliance officer or adviser. More questions are answered in the private AI answers hub.

What it costs

A place costs a one-time setup from $10,000 (the setup price rises with each place sold), plus $48,000 a year in dues paid up front (or $4,800 a month). Compute is billed to you at cost. Hands-on training and on-call support are extra, by the hour.

Questions

Will my saved prompts still work after a swap?

They may, and that is exactly what the pre-swap test checks. If a prompt behaves differently, that shows up before the new model goes live.

Can I ask for a specific new model by name?

You can ask, and it will be tested like any other candidate. If it does not beat the current model on your prompts or does not fit your licence needs, you stay put.

Does a swap change where my data lives?

No. The model is swapped, not your account. Your files stay in the cloud account you own.

What if a model I rely on is withdrawn by its maker?

A copy you already hold keeps running. Back up the files, and read the model's licence about continued use.

Do I pay extra each time the model is swapped?

Terms for model changes are part of what you agree with the provider up front. The questions to ask a provider page lists what to ask.

Related guides

Apply for a place

Takes thirty seconds. No payment now.

Adults only (18+). Private uncensored AI · Terms · Privacy · All guides