The Hugging Face steps run open models through Hugging Face's Inference Providers, on your company's own account. Each job is its own step, and each step lists only the models that do that job. Use them when an open model suits the work — a translation model for your languages, a small classifier that labels thousands of reviews cheaply.

Step What it does
Ask a model Answers a prompt with text or named fields, as the other AI steps do.
Classify text Gives the text a label — positive or negative, or labels you name.
Summarise text Shortens a long text.
Translate text Turns text into another language.
Create embeddings Turns text into vectors for search and matching.
Create an image Draws an image from a description.
Transcribe audio Turns a recording into text.

Who can do this

Workspace Admins and Editors, on every plan.

Before you start

A Hugging Face connection in Connections — see Connect Hugging Face.

Steps

Every Hugging Face step starts the same way:

  1. Select + where the step should go. In the step picker, open AI, then Hugging Face, and select the step.
  2. In Connection, choose the connection.
  3. Optional: in Model, choose a model. The list shows models that do this step's job and that Inference Providers can run, the most used first. Left empty, the step uses Bizomate's default for the job — a Llama model for Ask a model, Whisper for Transcribe audio, NLLB for Translate text, a model that takes any labels for Classify text. To use another, select Use a value and type its name as Hugging Face shows it — owner/name.

Then fill in the step's own settings:

  • Ask a model — Instructions (optional), Prompt, and Give me back: Fields (the default) or Text. For Fields, add a row per answer: the name, and what goes in it.
  • Classify text — Text, with values in {{ }} — {{ $json.review }}. Optional: Labels, the labels to choose from — delivery, billing, product — for a model that takes any labels; leave it empty for a model with its own, such as sentiment.
  • Summarise text — Text. Optional: Longest summary, in words.
  • Translate text — Text, To — for example German — and From, the language the text is in. The default model (NLLB) and mBART models need From; some models translate only one pair of languages, and their name says which.
  • Create embeddings — Text, or a list of texts, up to 512. Each text gets its own vector, in order.
  • Create an image — Prompt, and optional Leave out, Width and Height (in pixels, 64 to 2,048, in steps of 8).
  • Transcribe audio — under File, select Add a file and drag in the recording. See Pass files between steps.

Select Run this step, then check Output. Pin it while you build — see Pin a step's output.

Output

Step Keys
Ask a model each field by its name (Fields) or text (Text), and usage
Classify text label (the top one), score (0 to 1), labels (each with label and score, highest first), usage
Summarise text summary, usage
Translate text translation, usage
Create embeddings count, dimensions, embeddings (one list of numbers per text), first, usage
Create an image file (the image), contentType, usage
Transcribe audio text, segments (each with start, end and text, when the model gives them), usage

usage holds model, provider (which Inference Provider ran it) and seconds, with tokensIn and tokensOut where the model reports them. Output shows them as the usage strip.

Good to know

  • Each item is a separate call, and a separate cost on your Hugging Face account.
  • Which provider runs a model is Hugging Face's choice for Ask a model; the steps have no provider setting. The other six run on Hugging Face's own provider, the one that serves those jobs. usage.provider says which one ran.
  • Not every model can run. A model that no Inference Provider serves is left out of the list; typed by name, it fails with Hugging Face's reason.
  • A step may take up to five minutes; a model nobody has used for a while takes longest to start.
  • Results depend on the model. Try two or three on real examples before publishing; pin the best.

If something goes wrong

What the run says Why What to do
Hugging Face refused: … Check the connection's details on Connections. The token is wrong, revoked, or lacks Inference Providers. Edit the token in Hugging Face, or replace it.
Hugging Face refused: … The monthly allowance is used up, a rate limit, or the model cannot do this job. It is not tried again. Read Hugging Face's reason; choose a model from the list, check billing, or set the step to try again.
Hugging Face did not give back the fields asked for. Try again, or make the instructions plainer. The answer was not in the fields asked for. Make Instructions plainer, or the fields fewer.
NLLB models need the language the text is in. Choose it in From. From is empty, and the model needs it. Choose From.
NLLB models do not translate … The model has no code for that language. Choose another language, or another model.
File is empty. Choose a file from an earlier step. File holds no file in that run. Drag in a file from Data from earlier steps.
Could not reach Hugging Face: … / Hugging Face did not answer within 5 minutes. Hugging Face did not answer. Try again.