A model parameter is a learned weight or bias that encodes patterns from training data and helps determine how an LLM transforms one token sequence into the next. Models are often named by parameter count, such as 8B or 70B. section 1.2.1
Parameters differ from hyperparameters in how they are obtained and used:
- Parameters are learned during training; hyperparameters are configuration values chosen by practitioners, including inference-time settings that affect token generation. section 1.2.1
- Parameter count alone does not establish model quality or computational cost; mixture-of-experts architectures may activate only part of the model for an input. section 1.2.1








































