Model Version

An AI program that writes texts or recognizes images is not created once and then stays the same forever. Developers keep rebuilding it: they feed it new examples, change settings, or make it faster. Every state they capture and release along the way is a model version. So that these states can be distinguished, they get names or numbers, such as “4”, “4.1”, or a date like “2024-08-06”. Two versions of the same product can give different answers to the same question. The name on the box therefore says little unless you know which version is behind it.

Why the same question suddenly gets a different answer

Anyone using AI professionally builds workflows around it. A company, for example, might have it automatically sort customer emails. If the provider quietly switches to a new version in the background, the behavior changes from one day to the next. Sortings that used to be correct are suddenly wrong. That’s why large providers offer fixed versions that remain unchanged for a certain period of time.

The version information also matters for science. A study claiming that an AI passes a particular exam is worthless without a version specification. No one can verify the result if it remains unclear which version was tested. Serious publications therefore always state the exact designation along with the date.

A common misconception is that a higher number is better in every respect. This is often true, but not always. New versions are sometimes more cautious and reject more requests. Others are faster, but less accurate on difficult tasks. What counts as progress depends on what the model is needed for.

How a new version comes about

It starts with training: the model learns from vast amounts of text or images which patterns occur in them. This is the most expensive step and often takes weeks. This is usually followed by post-processing, in which humans evaluate which answers are good and which are undesirable. Only this finished package is then frozen and released as a version.

Not every new version means a complete retraining. Often an existing model is merely fine-tuned, for example to fix a weakness. Such small steps can be recognized by numbers like 3.5 instead of 4. A jump to a whole number usually stands for a fundamentally new architecture or significantly more training data.

A version also comes with a training data cutoff date. The model knows nothing about events after that point. If asked about something current, it may still make up an answer. Knowing the version helps to place such gaps in context.

Where these designations appear

In chat programs, the version is usually shown in a selection menu above the input field. There, you can often choose between a faster and a more thorough version. The faster one is cheaper to run, the more thorough one takes longer to answer. In free usage, you generally don’t get the latest version.

In news about technology companies, version names are a recurring topic. Stock prices react to whether an announced version meets expectations. What matters here is the difference between the product name and the actual version: a chat service can keep the same name for months while the model behind it changes several times. This is precisely why experts and regulators increasingly demand that providers disclose which version is currently in use.