What does the 7B or 70B in a model name mean?

When you see a name like Llama 2 7B or Llama 2 70B, the B stands for billion and the number counts the model’s parameters. So 7B is about 7 billion parameters and 70B is about 70 billion. Meta released Llama 2 in exactly these sizes: 7B, 13B, and 70B.^[1]

A parameter is a single adjustable number inside the model. During training, the model nudges these numbers up and down so that its outputs better match the text it is learning from. A finished model is, in a literal sense, just a very large list of these settled numbers. When people loosely call them “weights,” they mean the same thing.

The count is a rough measure of how much the model can store and how much computation each answer takes. It is not a quality score on its own. A 70B model has roughly ten times as many parameters as a 7B model, but that does not make it ten times better, and the relationship between count and capability is the subject of the next page.

A dense model’s B number is not always the whole story, either. Qwen3-235B-A22B, covered two pages over, carries two numbers because it is a mixture-of-experts model: 235B parameters stored, but only about 22B doing the work on any given token. Read against a plain 70B dense model, the MoE model stores more but can compute less per answer — the single B number that works cleanly for Llama 2 stops being enough once routing enters the picture.

The two-number convention exists because of how a mixture-of-experts model is built. A small router network, trained alongside everything else, looks at each token and decides which handful of “expert” sub-networks should process it. Every one of the 235B parameters has to be stored and ready, because any token might get routed to any expert, but only the roughly 22B belonging to whichever experts got picked for that specific token actually do any multiplying. That split is why the two numbers mean different things in practice: the larger one tells you the download and storage size, and the smaller one tells you roughly how much compute, and therefore how much time, each answer actually costs.