On October 7, Japanese AI company Preferred Networks (PFN) officially released its new-generation translation model PLaMo 3 Translate 31B. This fully Japan-developed model is built on the PLaMo 3 foundation model, with Japanese accounting for roughly 30% of its training data, and its average score across four translation benchmarks outperforms OpenAI's GPT-6.1 Sol and GPT-6 Astra.
Language Coverage Leaps: From 2 to 53 Languages
The most visible upgrade in PLaMo 3 Translate 31B is the expansion of its language support. The previous generation supported only two languages, Japanese and English; the new model jumps to 53 languages and introduces an online meeting translation mode covering 15 languages. This means the model can serve not only the traditional Japanese-English translation scenario, but also support real-time meeting communication in multilingual environments.
For a Japanese company, the quality of Japanese translation is its core competitiveness. PFN particularly emphasizes that the model is "proficient in Japanese while also handling other languages," with roughly 30% of its training data being Japanese — a proportion far higher than the share of Japanese in general-purpose large models, giving it a natural advantage in handling Japanese-specific honorific systems, context dependence, and elliptical expressions.
Performance Benchmarking: Average Score Across Four Benchmarks Surpasses the GPT-6 Series
PFN claims that PLaMo 3 Translate 31B's average score across four translation benchmarks surpasses those of OpenAI's two flagship models — GPT-6.1 Sol and GPT-6 Astra. Although the company has not released specific per-benchmark scores, making "surpassing the GPT-6 series" a core selling point shows PFN's full confidence in its Japanese translation capability.
It is worth noting that GPT-6.1 Sol is the flagship model OpenAI released in the second half of 2026, while GPT-6 Astra is its next-generation preview version. Being able to match or even surpass them in the vertical domain of translation shows that "specialized models" still possess advantages that general-purpose large models find hard to replace on specific language tasks.
Cost Advantage: Just 14 Yen per 100,000 Characters
In terms of pricing, PLaMo 3 Translate 31B's input cost is 14 yen per 100,000 characters (about 0.59 yuan RMB), a price PFN calls "significantly lower than competing products."
This cost advantage comes from two aspects: first, the model's parameter scale is kept at 31B, far below the hundreds of billions of parameters typical of general-purpose large models, so inference costs are inherently lower; second, its efficient Japanese processing capability reduces the number of tokens needed to handle the same content. For enterprise users who need to translate documents, contracts, or meeting minutes in large batches, this cost structure holds clear appeal.
A Microcosm of Japan's AI Self-Reliance
The release of PLaMo 3 Translate 31B is another milestone in Japan's push for a "sovereign AI" strategy. As one of Japan's most important domestic AI companies, PFN had previously launched the PLaMo series of foundation models. This release of a fully Japan-developed translation model is both a response to Japan's domestic multilingual translation needs and a step toward technological autonomy in the global AI supply chain.
In the global AI competitive landscape, Japan's choice differs from that of the United States, China, and Europe — it does not pursue an arms race in parameter scale, but instead focuses on deep optimization for Japanese-language scenarios, using "specialized, efficient, low-cost" as its differentiated competitive strategy.
The significance of PLaMo 3 Translate 31B lies in proving that "vertical specialized models" can still surpass general-purpose flagship models on specific language tasks. While most vendors fight at close quarters over general capabilities, PFN has chosen a narrower but deeper path — taking Japanese translation to the extreme. For a non-English-speaking country, this may be the most pragmatic AI self-reliance strategy: rather than competing with the world's strongest models across all dimensions, build an irreplaceable advantage on the battlefield of its own language.