Moegi: sharing service-specific language models internally
大規模テキストデータを活用したサービス特化型言語モデルの社内共有プロジェクト：Moegi
A · Strong evidence LY Corporation's language-processing team reports benchmarks built from real Yahoo! Shopping and Yahoo! Search tasks, comparing in-house pre-trained BERT models with public models and LLMs on accuracy and stated cost assumptions.
Summary based on the full text·Original: about 6 minutes to read
In short
A LINE Yahoo (LY Corporation) Tech Blog post (15 October 2026) on Moegi, an internal project that pre-trains BERT-class models on each service's own text (Yahoo! Shopping, Yahoo! Search and others) and shares them across product teams. On three classification tasks from real services, the 110-million-parameter in-house model beat two public Japanese BERT models on every task and matched or beat a zero-shot frontier LLM on two of three, at roughly one nine-hundredth to one seventeen-hundredth of its inference cost.
Key numbers
- 0.895 / 0.853 / 0.760Moegi bert-110m scores on Shopping query, Shopping product and Search query classification
- 0.909 / 0.848 / 0.749zero-shot gpt-5.6-sol (low reasoning effort) on the same three tasks
- 901.1x-1,740.7xinference cost of gpt-5.6-sol relative to the Moegi model across the three tasks
- 27.2x-58.4xinference cost of locally run gpt-oss-120b relative to Moegi, with lower accuracy (0.580-0.728)
As reported by the source.
Japanese·Web page·opens techblog.lycorp.co.jp
What it reports
- Public pre-trained models lack service-specific vocabulary, spelling variants and model numbers; separate pre-training by each product duplicates data, training, evaluation and maintenance work. Moegi centralises it and grants access to models on request through internal object storage.
- Benchmarks: Shopping query category (22 classes, accuracy), Shopping product category (22 classes, accuracy), Search query category (about 60 labels, multi-label micro-F1), built with data from user teams.
- Fine-tuned under the same conditions: cl-tohoku/bert-base-japanese-v3 scored 0.756 / 0.784 / 0.698, sbintuitions/modernbert-ja-130m 0.776 / 0.811 / 0.727, Moegi 0.895 / 0.853 / 0.760.
- Against prompted LLMs (zero-shot, low reasoning effort): gpt-oss-120b 0.708 / 0.728 / 0.580; gpt-5.6-sol 0.909 / 0.848 / 0.749. Moegi was higher on two tasks and close on the third.
- Cost relative to Moegi: gpt-oss-120b 27.3x / 27.2x / 58.4x; gpt-5.6-sol 1,740.7x / 1,566.6x / 901.1x.
- Moegi models are already used in several Yahoo! Shopping tasks (query intent, product attributes, recommendations, search suggestions).
How they got it
Same-condition fine-tuning for the BERT baselines; zero-shot prompts with category names and IDs for the LLMs. Costs are estimates: LLM cost from token counts and API or infrastructure prices (with prompt caching for gpt-5.6-sol); Moegi cost assumes an A30 GPU partition at up to 250 requests per second and 10% average utilisation.
Limits
- Three tasks, all classification; internal data, so results cannot be reproduced outside the company.
- LLMs were tested zero-shot without category definitions or examples; few-shot prompting or fine-tuning could narrow the gap.
- Cost ratios depend on the stated utilisation and pricing assumptions.
- Model sizes and training-data volumes for Moegi are not given.
Why it matters for companies
For high-volume, narrow tasks such as routing, tagging or classifying, a small model trained on your own text can match a frontier LLM at a fraction of the running cost. The organisational point matters too: one shared team that pre-trains and maintains models avoids every product team doing the same work. Firms with large proprietary text collections should test this before paying per-token prices for millions of classification calls.
Key passages in translation
Japanese original
各プロダクトが個別に事前学習済みモデルを作ると、データ整備、学習、評価、モデル管理などにかかる作業や管理の負担が重複します。
When each product builds its own pre-trained model, the work and management burden of data preparation, training, evaluation and model management is duplicated.
Japanese original
タスクごとの学習データでファインチューニングしたMoegiモデルは、zero-shotのLLMと比べ、3タスク中2タスクで高い評価スコアを示し、残る1タスクでも近い評価スコアとなりました。
Moegi models fine-tuned on each task's training data scored higher than the zero-shot LLM on two of the three tasks, and close on the remaining one.
Japanese original
Moegiモデルは、今回の3タスクで近い評価スコアを示したgpt-5.6-solと比べて、推論コストを約900分の1〜約1,700分の1に抑えられました。
Compared with gpt-5.6-sol, which scored similarly on these three tasks, the Moegi model cut inference cost to roughly 1/900 to 1/1,700.
Summary based on the full text. Read the original for the full argument and data.Read the original on techblog.lycorp.co.jp
- Tags
- Inference cost