# Researchers use GPT-5.4-nano to train cheaper home‑care scheduling model

> A team distilled live decisions from a frontier model into a small Qwen3‑8B model, cutting latency 2.5× and reducing cost while keeping accuracy.

Oossa · 2026-10-14 · https://oossa.com/en/researchers-use-gpt-5-4-nano-to-train-cheaper-home-care-scheduling-model

Keegan Wang and colleagues showed how to build a cheap scheduling assistant for home‑care agencies. They first collected real‑time yes/no safety decisions from the large GPT‑5.4‑nano model. Then they fine‑tuned a small 8‑billion‑parameter Qwen3 model on 11.6 k of those decisions. The small model agrees with the teacher 91.3% of the time, runs about 2.5 times faster, and costs a fraction of the original.

## The facts

- Paper published Oct 14 2026
- Qwen3‑8B fine‑tuned on 11.6 k decisions achieves 0.913 agreement with GPT‑5.4‑nano

## Why it matters

Home‑care agencies can deploy a low‑cost, faster model without sacrificing decision quality.

## Sources & references

1. [Iterate on an LLM, Serve on an SLM](https://arxiv.org/abs/2610.14576) – arXiv, 2026-10-14

Last updated: 2026-10-14