Skip to content

SYSTEM Cited by 1 source

GPT-4.1-nano

Definition

GPT-4.1-nano is OpenAI's smallest + cheapest member of the GPT-4.1 family, designed as a fine-tuning-friendly small model for classification, routing, and other low-latency tasks that don't need frontier reasoning. Yelp uses GPT-4.1-nano as the fine-tuning base for BAA's two pre-retrieval classifiers (Trust & Safety + Inquiry Type).

Role at Yelp (2026-03-27)

"We fine-tuned GPT-4.1-nano on a data set of a few thousands of questions-label pairs." (T&S classifier)

"Similarly to trust and safety we built our model by fine-tuning GPT-4.1-nano with ~7K samples." (Inquiry Type classifier)

"Fine tuned models: We fine-tuned smaller models using carefully curated training data generated from larger reasoning models to help with cost reduction for the question analysis phase. This in turn also helped us with the latency aspect as nano had quicker inference times compared to larger models." (Source: sources/2026-03-27-yelp-building-biz-ask-anything-from-prototype-to-product)

Load-bearing properties used at Yelp:

  • Small-model inference latency — "quicker inference times compared to larger models" — contributes to BAA's p75 < 3s target.
  • Fine-tuning on a few thousand labelled examples suffices for binary / multi-class classification tasks (T&S ~few- thousand samples; Inquiry Type ~7K samples).
  • Trained from teacher labels — Yelp's "carefully curated training data generated from larger reasoning models" is the teacher-student distillation discipline applied at the classifier altitude.

Comparison to adjacent OpenAI small-models

  • GPT-4o-mini — Yelp used this as the fine-tuned base for the 2025-02-04 query-understanding cache layer (segmentation + review-highlight expansion). GPT-4.1-nano is the newer small-model Yelp chose for the 2026-03-27 BAA classifier fleet.

Caveats

  • Stub page. OpenAI's disclosure of GPT-4.1-nano's architecture / parameter count / training data is limited; this stub captures only what's relevant to the Yelp BAA ingest.

Seen in

Last updated · 476 distilled / 1,218 read