Large language model: size is only one part of the recipe

A large language model (LLM) learns statistical structure in language at scale and uses context to produce token scores or representations. In a generative model, successive token choices can form text, code or structured outputs.

“Large” has no permanent parameter threshold. Architecture, training data, optimization and post-training all affect what a given parameter budget can do.

More parameters compete with more training

The Chinchilla study compared model sizes and training-token budgets under a fixed compute budget. It showed why allocating more computation to training a smaller model on more data can beat making the model larger without enough training.

That is a training tradeoff, not a universal serving rule. Repeated deployment can favour a smaller model trained more extensively because its per-request weight and compute costs are lower.

Total parameter count also differs from active parameters in a mixture-of-experts model. Context state and output length add resource demands beyond the weights. Compare models on the task and operating budget, not the size label alone.