What does 7B mean in a model name?
Seven billion parameters. The letter is just shorthand for billions, and the count is the single most useful number for telling whether a model will run on your machine.
7B means seven billion parameters, the learned numbers that carry everything a model knows. More parameters generally means more capability, and always means more memory, so the count is the number to check before downloading anything. A 7B to 8B model quantized to 4 bits fits in about 5 gigabytes, which is why that range runs comfortably on ordinary laptops and has become the practical sweet spot for local work.
Parameters are the learned numbers.
A model is not a file of rules, it is a file of parameters: billions of weights, each adjusted during training until the network's predictions fit the text it studied. Nothing about the model is stored anywhere else. When you download a 7B model you are downloading those seven billion numbers, and running the model is nothing more than multiplying through them. This is also why two models with the same count can behave differently: the parameter count says how big the memory is, never what was written into it.
The count sets the size.
Each parameter is a stored number, so seven billion parameters take seven billion numbers of space, compressed by quantization from their full 16-bit form down to 4 bits or less. The arithmetic is direct: a 7B model at 4 bits lands near 5 gigabytes, a 13B model near 10. That makes the count the practical gate for how much memory your machine needs, which is why hardware guides read like parameter guides.
Bigger is not automatically better.
A 70B model outperforms a 7B one on hard reasoning, but only if your hardware can actually hold it. A model too large for the memory either swaps to disk and crawls, or gets squeezed into a quantization so low that the quality advantage disappears. The honest way to choose is to start from what fits, then take the largest model that runs comfortably, since a smaller model answering quickly beats a larger one unusable slowly. For most consumer machines that lands in the 7B to 8B range, which is exactly where local research tools aim.
Why the letter varies in usefulness.
Model names also carry letters after the number, and they mean different things. A "B" for billion is a parameter count, while "MoE" in a name signals a mixture-of-experts model whose total count overstates what runs per token, since only a fraction of the experts activate at once. Version strings and fine-tune suffixes say nothing about size at all. So read the name in two passes: the number for memory planning, the rest for lineage. When a listing omits the count entirely, the file size and the quantization tag together tell you the same thing, because a 4.4 gigabyte GGUF at 4 bits is a parameter count in disguise.
The count is a floor, not the whole decision.
Two models with the same count can differ sharply in usefulness, because training quality varies more than size does. A well-trained 7B model beats a carelessly trained 13B on the questions people actually ask, which is why parameter counts rank below recency and training lineage when shortlisting models. The practical order of operations: check that the count fits your memory, then compare candidates on evidence you trust, and expect the same count to mean the same generation speed across models only roughly, since architecture shifts the constant. Treat 7B as the size class where good open models cluster, not as a quality grade.