n8n Hugging Face Inference Model nodeConfigure Hugging Face Inference Model in n8n.

Names the model that generates the completion. Everything else in the panel only shapes what that model does, so this is the first field to fill and the one to change when output quality moves in the wrong direction.

• Model: a text field holding the model identifier; set it once for a fixed model, or drive it with an expression like {{ $json.model }} when an earlier node decides which model to use for each item.

Points the node at a custom endpoint URL instead of the default inference service. Leave it empty in normal use and fill it only when the model lives somewhere you control.

• Custom Inference Endpoint: the custom endpoint URL, typed in full or pulled from an earlier node with {{ $json.endpoint }} when staging and production differ.

Positive values penalize new tokens based on how often they already appear in the text, which lowers the model's likelihood of repeating the same line verbatim. Reach for it when generated text starts circling back on itself.

• Frequency Penalty: a number; raise it when a long answer keeps recycling the same phrasing, leave it alone for short factual output where repetition is not a risk.

Caps how many tokens the completion may use. A token is a chunk of text, roughly a short word or a piece of one, and this setting is what decides whether an answer is a sentence or a page.

• Maximum Number of Tokens: a number bounded by the model's context length, which sits at 2048 tokens for most models, with the newest ones supporting considerably more.

Positive values penalize tokens based on whether they have appeared at all, which pushes the model toward new topics. Different lever from Frequency Penalty: that one fights repetition, this one fights staying on the same subject.

• Presence Penalty: a number; raise it when brainstorming output keeps orbiting one idea, keep it at rest when the task is to stay strictly on topic.

Controls randomness. Lowering it results in less random completions, and as the temperature approaches zero the model becomes deterministic and repetitive. It is the parameter to touch first when a workflow needs reproducible answers.

• Sampling Temperature: a number; near 0 for classification or extraction where the same input must give the same output, higher when the task is to write something new.

Controls how many token choices the model considers within the sampling operation when creating new text. Narrowing the pool keeps the output closer to the obvious continuation.

• Top K: a number setting the size of that candidate pool; a small value keeps generation tight, a large one lets less likely tokens back in.

Controls diversity through nucleus sampling: 0.5 means half of all likelihood-weighted options are considered. The official guidance is to alter this or the temperature, but not both at once.

• Top P: a number; lower it to ignore the less probable options, and pick between this and Sampling Temperature rather than tuning the two together.