Preferred Networks (PFN) launched plamo-3-translate, a large language model designed for translation, in its PLaMo Translation service on October 7, 2026. PFN says the model delivers world-class Japanese-English and multilingual translation at a fraction of the cost of existing translation and LLM APIs. Compared with plamo-2-translate, it used several hundred times more training tokens and computing power for data synthesis.

PFN trained the model specifically for translation on the third generation of its PLaMo language model. The 2025 plamo-2-translate focused mainly on Japanese-English and English-Japanese translation; although its Japanese was fluent, it sometimes paraphrased heavily, adding or omitting information, and had limitations in other languages. PFN rebuilt its translation datasets and used LLM-generated data to improve their quality. Development is ongoing, and PFN plans to publish the model weights on Hugging Face and release a blog post about its technical work.

In comparisons of the 8B model with its predecessor, the new model scored higher on several measures. Its PFTB English-to-Japanese score rose from 0.447 to 0.547, and its Japanese-to-English score from 0.602 to 0.672. Its English-to-Japanese score in PLaMo Translation's service evaluation rose from 0.430 to 0.571. On WMT24++, its English-to-Japanese COMET-22 score moved from 0.882 to 0.879, remaining nearly level. On FLORES+, chrF++ scores improved from 39.28 to 43.40 for Japanese-to-multilingual translation and from 29.90 to 31.04 for multilingual-to-Japanese translation. PFN says the new model also outperformed TranslateGemma 27B overall, despite that comparison model being more than three times the size of the 8B model.

The multilingual evaluation covered Arabic, Italian, Indonesian, English, Dutch, Korean, Spanish, Thai, Simplified and Traditional Chinese, German, French, Vietnamese and Russian. PFN also used Hungarian, Polish, Hindi, Persian, Bengali, Tagalog, Hebrew, Tamil, Burmese, Urdu, Malay, Nepali, Catalan, Malayalam, Bulgarian, Serbian, Croatian, Slovak, Slovenian, Sinhala, Azerbaijani, Georgian, Armenian, Uzbek, Lithuanian, Khmer, Latvian, Estonian, Basque, Cebuano, Belarusian, Telugu, Javanese, Kazakh, Marathi, Swahili, Gujarati and Punjabi in training. PFN says it also trained the model to some extent on translation between languages other than Japanese.

For its largest, 31B model, PFN compared performance and cost using public prices as of October 7, 2026. On Japanese-to-English translation, the model outscored nearly all comparison translation and LLM APIs at roughly one-thirtieth of their cost. The sole comparison system it did not outscore was Claude Fable 5.1, which performed better but cost about 85 times as much. PFN says the model also beat most APIs on multilingual translation at a fraction of their cost. Because PLaMo Translation does not publish API pricing, PFN used PLaMo API rates of 60 yen per 1 million input tokens and 250 yen per 1 million output tokens for the cost comparison.

In addition to evaluations against reference translations, PFN used Gemini 3.1 Pro to compare translations from plamo-3-translate 31B, DeepL and Google Translate (Translation LLM). It reversed the order of the systems and scored each sample twice; when the results differed, each side received half a win. Against DeepL, the 31B model's win rates were 87.00% on PFTB English-to-Japanese, 58.00% on PFTB Japanese-to-English, 73.82% on the service's English-to-Japanese evaluation and 59.17% on WMT24++ English-to-Japanese. Against Google Translate, the corresponding rates were 86.00%, 71.00%, 91.47% and 70.62%. The win rate exceeded 50% in every evaluation; PFN said the gap was particularly large in Japanese output.

PFN attributed the gains to Japanese making up close to 30% of PLaMo's training data, training focused specifically on translation and large, high-quality translation datasets. It also credited its in-house tokenizer (software that divides text into processing units) with improving cost efficiency. Across 50 English-to-Japanese and 50 Japanese-to-English PFTB samples, plamo-3-translate produced 447 Japanese tokens and 205 English tokens per 1,000 characters. Claude Fable 5.1 produced 893 Japanese tokens and 352 English tokens; GPT-6 Astra produced 744 Japanese tokens and 224 English tokens. PFN says the new model's Japanese output used about half as many tokens as Claude Fable 5.1. At the same token price and generation speed, PFN says, that means about half the cost for the same character count and about twice the generation speed per character.

PFN provided translation examples including polite, context-sensitive wording in business emails, natural phrasing in fiction and patent-style text beginning with “Claims:”. In a restaurant description, it translated “pints” as “beer” instead of retaining the unit name. PFN says these context-dependent strengths may not show up clearly in benchmarks.

PLaMo Translation offers

Preferred Networks (PFN) launched plamo-3-translate, a large language model designed for translation, in its PLaMo Translation service on October 7, 2026. PFN says the model delivers world-class Japanese-English and multilingual translation at a fraction of the cost of existing translation and LLM APIs. Compared with plamo-2-translate, it used several hundred times more training tokens and computing power for data synthesis.

PFN trained the model specifically for translation on the third generation of its PLaMo language model. The 2025 plamo-2-translate focused mainly on Japanese-English and English-Japanese translation; although its Japanese was fluent, it sometimes paraphrased heavily, adding or omitting information, and had limitations in other languages. PFN rebuilt its translation datasets and used LLM-generated data to improve their quality. Development is ongoing, and PFN plans to publish the model weights on Hugging Face and release a blog post about its technical work.

In comparisons of the 8B model with its predecessor, the new model scored higher on several measures. Its PFTB English-to-Japanese score rose from 0.447 to 0.547, and its Japanese-to-English score from 0.602 to 0.672. Its English-to-Japanese score in PLaMo Translation's service evaluation rose from 0.430 to 0.571. On WMT24++, its English-to-Japanese COMET-22 score moved from 0.882 to 0.879, remaining nearly level. On FLORES+, chrF++ scores improved from 39.28 to 43.40 for Japanese-to-multilingual translation and from 29.90 to 31.04 for multilingual-to-Japanese translation. PFN says the new model also outperformed TranslateGemma 27B overall, despite that comparison model being more than three times the size of the 8B model.

The multilingual evaluation covered Arabic, Italian, Indonesian, English, Dutch, Korean, Spanish, Thai, Simplified and Traditional Chinese, German, French, Vietnamese and Russian. PFN also used Hungarian, Polish, Hindi, Persian, Bengali, Tagalog, Hebrew, Tamil, Burmese, Urdu, Malay, Nepali, Catalan, Malayalam, Bulgarian, Serbian, Croatian, Slovak, Slovenian, Sinhala, Azerbaijani, Georgian, Armenian, Uzbek, Lithuanian, Khmer, Latvian, Estonian, Basque, Cebuano, Belarusian, Telugu, Javanese, Kazakh, Marathi, Swahili, Gujarati and Punjabi in training. PFN says it also trained the model to some extent on translation between languages other than Japanese.

For its largest, 31B model, PFN compared performance and cost using public prices as of October 7, 2026. On Japanese-to-English translation, the model outscored nearly all comparison translation and LLM APIs at roughly one-thirtieth of their cost. The sole comparison system it did not outscore was Claude Fable 5.1, which performed better but cost about 85 times as much. PFN says the model also beat most APIs on multilingual translation at a fraction of their cost. Because PLaMo Translation does not publish API pricing, PFN used PLaMo API rates of 60 yen per 1 million input tokens and 250 yen per 1 million output tokens for the cost comparison.

In addition to evaluations against reference translations, PFN used Gemini 3.1 Pro to compare translations from plamo-3-translate 31B, DeepL and Google Translate (Translation LLM). It reversed the order of the systems and scored each sample twice; when the results differed, each side received half a win. Against DeepL, the 31B model's win rates were 87.00% on PFTB English-to-Japanese, 58.00% on PFTB Japanese-to-English, 73.82% on the service's English-to-Japanese evaluation and 59.17% on WMT24++ English-to-Japanese. Against Google Translate, the corresponding rates were 86.00%, 71.00%, 91.47% and 70.62%. The win rate exceeded 50% in every evaluation; PFN said the gap was particularly large in Japanese output.

PFN attributed the gains to Japanese making up close to 30% of PLaMo's training data, training focused specifically on translation and large, high-quality translation datasets. It also credited its in-house tokenizer (software that divides text into processing units) with improving cost efficiency. Across 50 English-to-Japanese and 50 Japanese-to-English PFTB samples, plamo-3-translate produced 447 Japanese tokens and 205 English tokens per 1,000 characters. Claude Fable 5.1 produced 893 Japanese tokens and 352 English tokens; GPT-6 Astra produced 744 Japanese tokens and 224 English tokens. PFN says the new model's Japanese output used about half as many tokens as Claude Fable 5.1. At the same token price and generation speed, PFN says, that means about half the cost for the same character count and about twice the generation speed per character.

PFN provided translation examples including polite, context-sensitive wording in business emails, natural phrasing in fiction and patent-style text beginning with “Claims:”. In a restaurant description, it translated “pints” as “beer” instead of retaining the unit name. PFN says these context-dependent strengths may not show up clearly in benchmarks.

PLaMo Translation offers