# What are frontier models?

Official page: https://aitje.com/en/knowledge/wat-zijn-frontier-modellen

> Frontier models are at the cutting edge of what AI is currently capable of. What does that say about quality, and when do you really need such a model?

### In brief

A frontier model is an AI model that is currently among the most advanced and capable models available. ‘Frontier’ means boundary or vanguard. It is not a fixed quality mark, nor is it a specific architecture: what is at the forefront today may be overtaken later. These are often versatile models capable of handling multiple types of complex tasks.

### What makes a model a ‘frontier’ model?

This concerns capabilities such as following complex instructions, solving problems, writing code or combining text and images. A model may excel at one task but be less effective at another. A high overall benchmark score therefore does not necessarily indicate how well it understands your Dutch documents or carries out your business processes.

### ‘Frontier’, ‘open’ and ‘local’ are distinct concepts

‘Frontier’ describes the capabilities. ‘Open’ or ‘closed’ refers to access and usage rights. Whether locally or via an external API refers to where the model is run. These characteristics do not automatically coincide. A local model is not necessarily simple, and a ‘Frontier’ model does not necessarily have to be closed. The available hardware and licence partly determine what is practically possible.

### A practical example

Suppose you want to categorise incoming messages into five categories and then work out the tricky exceptions. For the first step, you can test a small model. Only the complex messages are sent to a more powerful model. This is a possible workflow, not a proven performance claim. You compare the results against the same examples and clear acceptance criteria.

### When does a more powerful model make sense?

A more robust model can add value if errors are costly or the task requires a great deal of context and judgement. It may also mean longer waiting times or higher costs. You should therefore assess how many usable results you get, how much correction work is required and how data is processed. A model name or popularity alone is not a selection criterion.

### AITJE’s perspective on this

Start with the task and test different models on the same input. For each component, choose sufficient quality, with appropriate costs and control. Sometimes a frontier API is suitable, sometimes a model running on your own hardware, and sometimes a combination. By keeping that choice separate from the rest of your workflow, you can more easily deploy a different model later on.

## Sources

- [NCSC — What is meant by frontier AI?](https://www.ncsc.gov.uk/section/advice-guidance/all-topics/frontier-ai)
- [RouteLLM — Research into model routing and the trade-off between quality and cost](https://arxiv.org/abs/2406.18665)

## Contact

[Contact AITJE](https://aitje.com/en/contact) — contact@aitje.com