Can I move between sizes without retesting?

You can move without re-engineering, and testing is exactly what you should not skip. The interface, the billing and the general behaviour carry over, which is the point of a family, while the specific work that was marginal on the larger member is what fails first on the smaller one. Test the hardest recurring case rather than a representative one.

Does staying inside one family lock me in?

Less than it feels, because interfaces across vendors have converged enough that switching is usually a configuration change rather than a rebuild. What genuinely accumulates is tuning to one family's habits: the phrasings that work, the shapes it returns reliably. That is real and it is smaller than the effort most teams imagine.

Why do vendors keep renaming their families?

Because the name carries the marketing weight of a generation, so it moves when the generation does. The practical consequence is that a document naming a family ages faster than one naming a specific member and generation, which is why anything load-bearing should record the exact member it was tested against rather than the family it belonged to.

Glossary

Model family

A model family is a set of releases from one maker sharing a name and an underlying design, usually offered in several sizes so cost can be traded against capability without changing supplier.

In plain terms

Several versions of the same thing from one maker, sold under one name in different sizes. The big one costs more and handles harder work, the small one is cheap and fast, and they are built to behave similarly enough that moving between them is a setting rather than a project. The name is a marketing decision, so it promises less than it appears to.

Why it matters

Because the shared name invites an assumption the vendor never made. A family markets itself as one thing offered at several price points, which makes swapping sizes feel like adjusting a dial, and the differences between the largest and smallest members are frequently wider than the differences between two rival products. Testing across sizes matters more than testing across vendors.

How it works

The name signals a common design and a common release moment, not equivalent behaviour. Members usually share an architecture, a training approach and a generation, which is what makes their outputs feel related and their interfaces interchangeable, and none of that guarantees that the small one handles your task the way the large one did. Similar is doing a lot of work in that sentence.

Sizes exist so cost can be traded against capability without changing supplier, and that is the family's real product. Routine high-volume work goes to a smaller member at a fraction of the price while the hard cases go to the largest, which is the same two-tier arrangement the frontier discussion arrives at, expressed inside one vendor's line instead of across two. Staying in one family keeps the interface and the billing unchanged.

Families now split on openness as well as size, and that split is the one that surprises people. A single line can offer downloadable members you run yourself alongside a flagship that is available only as a hosted service, so the family name tells you nothing about whether you can hold the weights. One vendor's own page states exactly this: the current flagship is closed and only the other members can be self-hosted.

Membership changes underneath the name, which makes a family a moving target rather than a catalogue. New members arrive, older ones are retired on the vendor's schedule, and the same name can mean something materially different a year apart, so anything written down should record which member and which generation rather than the family. A pinned member is a decision; a family name is a direction.

What the shared name does and does not tell you

The reason a family name carries more weight than it should is that it is doing genuine work in one direction and none at all in the other. Told that two products belong to the same family, you can reasonably expect the code that calls one to call the other, the invoice to arrive from the same place, and the general character of the output to feel continuous. That is worth something and it is why families exist as a commercial shape. What the name cannot carry is any promise about the specific work you need done, because the distance between members is set by their size rather than by their branding, and your task sits wherever it sits on that curve. The habit that prevents the surprise is small: when a document or a decision references a family, ask which member was actually tested, and if the answer is that the family was tested, nothing was.

Seen in the wild

A line whose downloadable members are the ones you self-host, while its current flagship is available only through a hosted interface.
Qwen
A downloadable family widely supported across tooling, with no first-party hosted tier at all, so you run the weights or use somebody else's hosting.
Llama (Meta)
An assistant offering slower, more capable members for multi-step work alongside faster ones for everyday questions, under one subscription.
ChatGPT

Common misconceptions

People assume

Members of a family behave the same, just faster or slower.

In fact

They share a design and diverge on exactly the work that is hard for you. The gap between the largest and smallest members of one family is often wider than the gap between two vendors' comparable sizes, so a task that survives the swap in testing is the only evidence that it survives the swap.

People assume

A family name tells you how it can be deployed.

In fact

It tells you who made it and roughly when. Openness now varies within a single line, so a family can contain both weights you download and a flagship you can only reach as a service, and reading the family name as a deployment answer is how teams discover late that the member they wanted is the one they cannot host.

Telling them apart

Model family vs Foundation model

Model family

A vendor's set of related releases, sold in sizes.

Foundation model

A general base others adapt for many jobs.

One is a product line and the other is a role, so a family usually contains foundation models and the words are not describing the same kind of thing at all.

Questions

Can I move between sizes without retesting?: You can move without re-engineering, and testing is exactly what you should not skip. The interface, the billing and the general behaviour carry over, which is the point of a family, while the specific work that was marginal on the larger member is what fails first on the smaller one. Test the hardest recurring case rather than a representative one.
Does staying inside one family lock me in?: Less than it feels, because interfaces across vendors have converged enough that switching is usually a configuration change rather than a rebuild. What genuinely accumulates is tuning to one family's habits: the phrasings that work, the shapes it returns reliably. That is real and it is smaller than the effort most teams imagine.
Why do vendors keep renaming their families?: Because the name carries the marketing weight of a generation, so it moves when the generation does. The practical consequence is that a document naming a family ages faster than one naming a specific member and generation, which is why anything load-bearing should record the exact member it was tested against rather than the family it belonged to.

Key takeaways

The shared name promises a common design, never equivalent behaviour.
Sizes exist to trade cost against capability without changing supplier.
Openness now varies inside one family, so the name says nothing about deployment.
Record the member and generation; a family name is a direction, not a decision.

Tools that use this

Qwen
Downloadable members self-hosted; the current flagship hosted only.
Llama (Meta)
A downloadable family with no first-party hosted tier.
ChatGPT
Slower capable members and faster ones under one subscription.

Last checked July 2026

All glossary terms