What is the difference between a frontier model and a local model?

There are two ways to use a large language model, and the difference is mostly about where the model physically runs.

A frontier model is the newest, most capable model a vendor offers. You do not get a copy of it. It runs on the vendor’s servers, and you send it requests over an API, paying for each one. The vendor controls the model, updates it, and can change or retire it. You never see the weights, the design, or the training data; from your side it is a sealed box you talk to.^[1]

A local model is one whose weights (the trained numbers that make it work) are published for download. You pull those weights onto your own laptop, workstation, or server and run them there. Once downloaded, the model is yours to run offline, inspect, and fine-tune for a specific task.^[1] These are usually called open-weight models, and they are the practical option for running AI on your own machine.

The two are not rivals so much as different tools. Frontier models give you the highest capability with nothing to set up. Local models give you control, predictable cost, and data that never leaves your hardware. Which one fits depends on the job, and that comes down to three trade-offs: capability, cost, and privacy.

The line is not always as sharp in practice as the test makes it sound. A maker can put a downloadable model behind a paid hosted API at the same time (most open-weight families do exactly this, for anyone who would rather not run the hardware themselves), and a company can self-host an open-weight model on rented cloud servers, which is neither a pure frontier setup nor a laptop in the corner. The download test tells you what kind of model it is; it does not tell you how any particular person chose to run it.