Parameters
Also called model parameters, weights, model size
The internal numbers a model learns during training. Model size, like "70B", counts them in billions.
Parameters are the adjustable numbers inside a neural network, mainly the weights connecting its layers. Training changes them, a tiny bit at a time, so the model gets better at predicting text. Once training ends, they are fixed, and everything the model "knows" is encoded in them. When a model is described as 8B or 70B, that means about 8 or 70 billion parameters.
Parameter count is a rough guide to capability and cost. More parameters generally let a model store more knowledge and handle harder tasks, but they also need more memory and computation to run. A 70B model stored at 16-bit precision needs about 140 GB of memory just for its weights, which is why quantization, storing each parameter in fewer bits, is so common for running models locally.
Size is not everything. Training data quality, training length, fine-tuning and architecture choices matter a great deal. Modern small models often beat much larger models from a few years earlier. Some architectures, called mixture of experts, have many parameters but use only a fraction of them for each token, which makes them cheaper to run than their total size suggests.
Example
A laptop with 16 GB of memory can run an 8B model quantised to 4 bits (about 5 GB) comfortably. The same model at full 16-bit precision (about 16 GB) would not fit alongside the operating system.
Common mistake
Using "parameters" to mean API settings such as temperature. In everyday developer talk both are called parameters, but model parameters are the learned weights inside the model; temperature, top-p and max tokens are settings you choose per request.