AI Systems Research guide

Frontier, workhorse, or local: three classes, not a religious war

The forum fight is cloud versus local. The operator question is which class owns this task.

Three workbenches in one workshop: marble precision, scuffed steel bulk, oak tools behind glass
Three classes in one room. The job picks the bench.

Direct answer

Frontier, workhorse, and local are three classes of language-model work from The Model Portfolio. Frontier is the scarce, expensive tier, rationed for tasks that fail on anything cheaper. Workhorse is the default-capable tier where volume is meant to live. Local is a placement decision: the model runs on infrastructure you control because privacy, latency, sovereignty, or volume says it must. Frontier and workhorse are capability tiers. Local is about where the model runs, and a local model can serve a workhorse route or even a frontier route if your own evaluation says so. The book also names a fourth archetype, specialist, for narrow domains. A private, high-volume, easy task is not a frontier job. The Model Portfolio is live. This page is the split, not the routing policy.

Search “local LLM vs API” and you get a forum fight. One camp treats the cloud as betrayal. The other treats a local model as a hobby. Neither is an operating split.

Frontier, workhorse, or local? Three classes. The task picks the class. The brand does not.

Frontier is the scarce tier. Deep reasoning, messy ambiguity, long-context synthesis, jobs where a cheaper miss costs more than the token bill. You ration it. The moment frontier becomes the default destination, the invoice starts explaining your architecture.

Workhorse is the default-capable tier. Bulk extraction, classification, summaries, drafts a human will actually read. This is where volume is meant to live. If the class is empty, every task inflates to frontier.

Local is a placement, not a capability rank. A model is local when control of the infrastructure it runs on is part of its mandate: on-premise, a private VPC, an edge device, a developer laptop. Privacy class “never leaves”, a latency budget inside the room, or a volume that would turn the API into a tax. A local model can clear the bar of a workhorse route, or a frontier route, if your own evaluation says so.

Not a cloud religion, not a leaderboard row

“Open versus closed” is a procurement argument. A closed workhorse and an open workhorse can share a mandate.

A leaderboard rank does not know the privacy class of your prompt. If the prompt carries customer data you promised not to export, the class is local or the call does not happen.

The Model Portfolio names a fourth archetype, the specialist: a model fine-tuned or built for one narrow domain such as code, legal extraction, or a single language pair. It earns a place only when a general model at any tier fails a well-defined task, and the task is frequent enough to pay for the maintenance. This page keeps to the three most people search for.

How to choose this week

Write the task in one sentence. Then score three things, in this order.

Privacy. May this text leave the building? If not, local. Stop scoring.

Difficulty. Does a cheaper miss cost more than the tokens you save? If yes, frontier. If no, workhorse.

Volume. Will it run thousands of times a day? Then workhorse or local, never frontier by default.

Write the answer next to the task, so next week’s hire does not reverse it in a chat window. A vendor SKU is not a class. You assign the class, and the same hosted model can be workhorse on Tuesday and a forbidden privacy miss on Wednesday.

Two pages

What is model routing? is the policy that uses these classes. This page is the split.

The Model Portfolio is live. You do not need the hardcover to stop sending private bulk to the flagship.

Cite this:Frontier, workhorse, or local: three classes, not a religious war.Len P. van der Hof. https://lenvanderhof.com/en/blog/frontier-vs-workhorse-vs-local/ ·

Terminology

Sources

  1. ROUTE (glossary)
  2. What is model routing?
  3. The Model Portfolio

Further reading

Markdown for LLMs