What is a large language model (LLM)?

Why it matters

Most LLMs are built on transformer networks and can recognise, summarise, translate, predict and generate content. A model's knowledge is held in its parameters, the numerical values learned during training. In a dense model, every parameter takes part in each step. In a mixture-of-experts model, a routing network picks a subset of parameters for each token, so the active count is lower than the total.

Model size is one of the main drivers of AI energy use. Larger models need more compute, and so more electricity, for every token they process. Of all the choices a team makes about AI, picking the model makes the biggest difference to its energy and cost.

An example

In Greenpixie's data from October 2026, one large OpenAI model uses about 49 times the energy of a smaller model from the same provider for every output token, at around 100 times the list price. Smaller models suit classification, extraction and short answers. The largest models are best kept for work that needs them.

How Greenpixie measures it

Greenpixie benchmarked 32 open-weights LLMs, from 1.7 billion to 1 trillion total parameters, covering both dense and mixture-of-experts designs. Regression models predict energy per token from two features only, being active parameters and the memory the weights take up on the GPU. GPX AI applies this to over 500 models.

Proprietary models do not publish their size, so each is assigned a bounded parameter band based on its provider's tier naming. Their energy is reported as a range.