Japan's well-known enterprise Preferred Networks (PFN) officially launched a new translation model - PLaMo3Translate31B, yesterday local time. This model not only demonstrates a high level of professionalism in Japanese translation, but also takes into account the translation needs of other languages, making it a completely independently developed large model product by Japan. In terms of specific training data composition, Japanese accounts for about 30%.

image.png

From a technical architecture perspective, the model is built upon the PLaMo3 base model. Compared to the previous generation product that only supported two languages, Japanese and English, PLaMo3Translate31B has achieved a leap in the number of supported languages, reaching 53, and has innovatively introduced an online meeting translation mode covering 15 languages.

In terms of performance and cost, PFN officials stated that PLaMo3Translate31B outperformed OpenAI's GPT-6.1Sol and GPT-6Astra in the average score of four authoritative translation benchmark tests. Thanks to its efficient processing mechanism for Japanese, the operational cost of this model is also impressive, with a cost of only 14 yen (about 0.59 Chinese yuan) per 100,000 characters input, significantly lower than similar products in the market.