Do bigger and smaller AI models in the same family write noticeably differently, in ways you could detect?

This explores whether models of different sizes within one family leave different fingerprints on their outputs. The collection doesn't directly study telling a small model's text from its bigger sibling's. It does have several findings that point the same way: size changes how a model distributes its choices, not just how accurate it is, so outputs can differ in recognizable ways.

The clearest clue is about diversity. For synthetic data generation, models around 500M parameters produce more unique outputs per sample than larger ones, because bigger models concentrate probability on a few preferred answers Why aren't bigger models better for generating diverse outputs?. A small model's outputs are spread out and varied, while a large model's are narrower and more repetitive. Run each many times on the same prompt and that difference in spread is itself a signal. Size also changes behavior. Larger and instruction-tuned models follow beliefs a user states in the prompt less often and lean more on what they learned in training Do larger models follow stated beliefs less often?. On the same prompt, two siblings can give different answers because they weigh the context differently.

The less obvious finding is about post-training. When RL is applied to models that were pretrained on mixed formats, it quickly amplifies one format and suppresses the rest. Which format wins depends on model scale, not on which one performs best Does RL training collapse format diversity in pretrained models?. So two sizes in one family could settle on different 'house styles' even with the same training recipe. Because that choice happens inside the provider's pipeline, outsiders mostly can't see it. Preference tuning makes this messier, because its effect on diversity flips by domain: it narrows code but can widen creative writing Does preference tuning always reduce diversity the same way?. How distinguishable two models are may depend on what you ask them to write.

For comparison, telling different families apart is easy. ChatGPT and Claude can be identified 85.5% of the time from opening phrases and markdown habits alone What specific phrases reveal which model wrote a response?. Those habits mostly come from each company's post-training, which siblings in one family share. So the differences between sizes are probably subtler: less about signature phrases and more about how varied the outputs are and how closely they follow the context. Matching benchmark scores don't settle the question either. Models with identical accuracy can be organized very differently inside Can models be smart without organized internal structure?, and a well-trained 3B model can match much larger systems on verifiable reasoning Can small models match frontier reasoning without massive scale?. Two models can score the same and still behave differently.

In short, to tell siblings apart, look at the spread of their outputs rather than at word choice. Bigger models tend to be more predictable, more anchored to what they learned in training, and possibly locked into a different format. Nobody here has yet built a detector for model size within a family, which makes it an open gap in the collection.