# Fine‑tuned 0.8B model trims grammar annotation costs 16‑fold

> A small language model, Qwen3.5 0.8B, now tracks English learners’ grammar mastery cheaper and more accurately than GPT‑5 prompts.

Oossa · 2026-10-09 · https://oossa.com/en/fine-tuned-0-8b-model-trims-grammar-annotation-costs-16-fold

A language‑learning platform has rolled out a tiny AI model to label grammar concepts in student‑tutor conversations. The model, a 0.8 billion‑parameter version of Qwen3.5, was fine‑tuned on teacher‑created examples and now runs as the engine behind a new grammar‑mastery tracker for all English learners on the service. The company says the system is 16 times cheaper to serve than using large frontier models such as GPT‑5.4 or GPT‑5.6 with prompts.

## How the model was built and tested

The team filtered and rebalanced teacher‑generated supervision data, then added an adapter layer – a small set of extra weights – that encodes the annotation rules. This lets the 0.8 B model work with a short, matched prompt instead of long instructions. In two human‑curated benchmark tests, the deployed model and a 4 B reference model beat prompted GPT‑5.4 and GPT‑5.6 on both precision and recall, using increasingly strict matching criteria (concept, evidence span, correctness).

## Impact on learners and business

An online A/B experiment showed the new tracker lifted learner engagement by 15.8 percent. Scheduled lesson hours rose 2.1 percent and gross merchandise value from new lessons grew 13.2 percent, according to the platform’s internal metrics.

## The facts

- The fine‑tuned model is a 0.8 B‑parameter Qwen3.5 SLM.
- Serving cost is reduced by roughly 16 times compared with prompted GPT‑5 models.
- On two benchmarks the 0.8 B model outperformed GPT‑5.4 and GPT‑5.6 in precision and recall.
- Learner engagement increased by 15.8 % in a feature‑level online experiment.
- Scheduled lesson hours rose 2.1 % and new‑lesson revenue grew 13.2 %.

## Why it matters

For English learners, the cheaper AI means the platform can offer continuous, automated grammar feedback without raising prices. For the company, lower compute costs and higher engagement translate into more lesson bookings and revenue.

## Sources & references

1. [Grammar Concept Annotation at Scale: Deployed Fine-Tuned Small Language Models Outperform Prompted Frontier Models](https://arxiv.org/abs/2610.10827) – arXiv, 2026-10-09

Last updated: 2026-10-09