Preferred Networks Releases PLaMo 3 Translate 31B Model for Japanese Translation
Japanese AI company Preferred Networks (PFN) has released PLaMo 3 Translate 31B, a translation model that supports 53 languages with a strong focus on Japanese. Built on the PLaMo 3 base architecture, the model incorporates roughly 30% Japanese training data and introduces an online meeting translation feature covering 15 languages. PFN reports that PLaMo 3 Translate 31B outperforms top frontier models like OpenAI's GPT models across four translation benchmark tests while operating at a fraction of the cost. This highlights how language-optimized specialized models can deliver superior localized accuracy and lower tokenization overhead compared to generalist large language models. The release expands PLaMo's language coverage from just 2 languages (Japanese and English) in the previous generation to 53 languages. It offers high efficiency for Japanese character processing, priced at only 14 Japanese yen (approximately 0.59 RMB) per 100,000 input characters.
## BACKGROUND
Preferred Networks (PFN) is a Japanese AI startup focused on deep learning innovation, vertically integrating hardware, cloud infrastructure, and large language models like the PLaMo series. Dedicated translation models fine-tuned with high-quality regional data often outperform generic global models in preserving local nuances and technical terms.