← All articles

ARTEMIS LAB / BLOG

Hugging Face: who is running your AI?

Artemis Lab3 min read

OpenRouter now routes language-model requests on Hugging Face. What to check before trusting the service with a task.

Your assistant can keep using the same AI model while another company processes its requests. On November 16, Hugging Face announced that its language-model requests now go through OpenRouter. Existing users should check one migration detail: custom provider keys are no longer supported for these models, with billing handled through the Hugging Face account. Official announcement.

The model and the machine running it

Think of a recipe prepared in several kitchens. The recipe represents the model; the kitchen represents the provider running it. Picking a recipe does not, by itself, tell you who prepares your meal.

“Routing” is the step that picks the kitchen. The documentation describes selection based on price, speed and reliability, with another attempt elsewhere if a provider fails. An explicitly pinned provider is an exception: it is never automatically replaced. Routing rules.

Our interpretation: this flexibility is especially worth evaluating when an application needs a response to keep working. It does not, on its own, establish that the response will be correct. Getting an answer and being able to trust it remain separate questions.

Start with a task you can check

Consider a fictional example: sorting information requests into “opening hours,” “delivery” and “other.” Before connecting the tool to your inbox, write a small set of invented messages. For each one, record the expected category yourself, including cases where more than one answer would be acceptable.

Keep the instructions unchanged during the trial. Record the result, the delay and mistakes that need human correction. Include an ambiguous message: a useful AI should also be able to signal that information is missing.

This is a method proposed by Artemis Lab, not a test we have conducted. It addresses a specific question: does the service produce something usable for this task? An impressive demonstration on a single question cannot settle that.

Decide who may receive requests

Hugging Face lets you disable providers. Requests billed to an organization follow that organization’s settings. Settings documentation.

Before trying internal documents, have your team write down the providers it accepts. Check the settings of the account paying the bill, then start with fictional data. Choosing a model should not replace the separate decision about where information goes.

Also decide who will review responses and handle mistakes. For our fictional sorting task, you could require human approval before any reply is sent. That gives the first trial a clear boundary without handing over the whole customer relationship.

Check the migration before expanding

The announcement describes a service change; it is not an independent evaluation of your application. We have not measured any speed improvement or savings. If you already used a provider key, first check one complete request and the account being billed.

Next, compare the cost of the completed task, including corrections, with your current method. Keep a way to return to that method if the trial fails. The first useful outcome is a documented decision about a small task, with someone responsible who can explain where requests go.

IAActualités