Beyond 'Strong and Powerful Morning': An Experiment Log on TranslateGemma as a Replacement for GPT-4o-mini

Musinsa began reexamining global product review translation from an operational perspective. The core point was that issues like residual Korean text, mistranslations, failure to preserve brand names, and colloquial expression handling — not just whether the translation read naturally — quickly became operational risks.

Previously, they used GPT-4o-mini, but as the number of reviews grew, API translation costs grew along with it, and limitations emerged in consistently handling fashion-domain-specific terminology and tone. In particular, as cases accumulated where 'oriteol paeding' (duck-down padding jacket) was translated almost as a phonetic transliteration, or where kk (a Korean laughter marker, similar to 'lol') remained in English/Japanese sentences, they began reviewing on-premise translation models.

After encountering TranslateGemma, announced on January 15, 2026, they began experimenting immediately after the announcement. The experiment covered 11 Korean product reviews, translating sentences averaging 583 characters from Korean→Japanese and Korean→English for comparison. The models compared were TranslateGemma 27B(Q6), GPT-4o-mini, and gpt-oss:20b.

Evaluation was conducted on a 100-point scale.

• Quantitative 50 points: untranslated text (residual Korean), errors, stability, speed/operability
• Qualitative 50 points: special expressions/colloquialisms, terminology/brand/proper nouns, naturalness

As a result, TranslateGemma 27B(Q6) showed the most operationally friendly translations. Meaning-based translations using expressions common in Japanese, such as 'down jacket,' 'sweat,' and イエベ, came through well, and sentence flow didn't significantly damage the review tone. In contrast, GPT-4o-mini notably showed cases of transliterating words by pronunciation, producing context-inappropriate expressions, or leaving untranslated strings.

The key point wasn't just translation quality. What mattered was that the 27B model with Q6 quantization applied lowered cost and resource usage while finding a point where fewer problems occurred in actual operation. The author concludes from this experiment that even domain developers can directly verify translation models to build sufficiently operational alternatives.