Can you tell two AI models apart more easily by how they're built than by how big they are?

This explores whether a model's design leaves a clearer trace than its size does. One caveat first: the library has no papers on fingerprinting, meaning working out which model wrote a given output. What it does have is evidence on a related question: when do architecture and scale cause visible differences in behavior, and when do they stay hidden? Taken together, that evidence suggests that architecture often shows up in *how* a model fails, while size shows up mostly in *how well* it does.

The best example is the instruction-overload research. When models get more and more instructions at once, they break down in different ways. Small models get worse at a steady rate. Mid-sized models collapse faster and faster. Reasoning models hold steady up to about 150 instructions and then fall off a cliff How does instruction density affect model performance?. Model type changes the *shape* of the failure curve, not just its height. That is a better signal than an accuracy score, because two models can tie on the score and still have curves that look nothing alike. Size differences, by contrast, can show up in unexpected places. Smaller models around 500M parameters produce more varied outputs, because larger models pile probability onto the answers they prefer Why aren't bigger models better for generating diverse outputs?. So a narrower spread of outputs may hint at scale more than any capability test would.

Architecture also matters much more than the older scaling story assumed. At small scale, deep-and-thin models beat balanced ones of the same size by a noticeable margin Does depth matter more than width for tiny language models?. Scaling laws that include design choices like hidden size or the split between attention and feed-forward layers can predict large differences in speed and accuracy under the same training budget Can architecture choices improve inference efficiency without sacrificing accuracy?. Mixture-of-Experts can remove the competition between vision and language in multimodal models. Dense models of any size can't do that, because the problem comes from how capacity is allocated, not from how much capacity there is Can we solve modality competition through architectural design?.

The problem is that standard benchmarks can hide all of this. Models can score perfectly while their internal organization is badly fractured, and the damage only shows up under perturbation or distribution shift Can models be smart without organized internal structure?. Size differences can be hidden too: a smaller model given more thinking time can match a bigger one on hard prompts Can inference compute replace scaling up model size?. Targeted DPO training can bring small models up to large-model accuracy on function calling Can small models match large models on function calling?. The code wrapped around a model can lift a fixed model's scores with no changes to its weights Can execution harnesses lift model performance without retuning weights?. So in a deployed system, what you see may reflect the scaffolding as much as the model underneath.

The takeaway you might not expect is this: if you want to tell models apart, don't compare their average scores. Push them until they break, and look at the shape of the breakdown and how varied their outputs are. Architecture tends to show up in failure patterns, and size in output diversity. Either one can be hidden by extra inference compute, training, or a strong harness.