The well-known Japanese company Preferred Networks (PFN) officially released a brand-new translation model yesterday local time — PLaMo3Translate31B. The model not only demonstrates an extremely high level of professionalism in Japanese translation but also addresses translation needs in other languages, making it a large model product entirely self-developed in Japan. In terms of the specific composition of the training data, Japanese accounts for approximately 30%.
In terms of technical architecture, the model is built on the PLaMo3 base model. Compared with the previous generation product, which supported only 2 languages, Japanese and English, PLaMo3Translate31B achieved a leap in the number of supported languages, reaching 53 at once, and innovatively introduced an online meeting translation mode covering 15 languages.
In terms of performance and cost, PFN officially stated that PLaMo3Translate31B's average scores across four authoritative translation benchmarks surpass OpenAI's GPT-6.1Sol and GPT-6Astra. Thanks to its efficient processing mechanism for Japanese, the model's operating cost performance is equally impressive: the cost per 100,000 input characters is only 14 yen (approximately 0.59 RMB), significantly lower than comparable competitors on the market.