Preferred Networks opens PLaMo 3 Translate, offline translation models from 0.8B to 31B

All four models are on PFN's Hugging Face page today[2][3]. They were first developed for PFN's PLaMo Translate service and are based on PLaMo 3 NICT, a pre-trained model PFN built with Japan's National Institute of Information and Communications Technology[1]. They cover 53 languages, including English and Japanese[1].

The spread of sizes is the point. The smaller models are meant for phones, laptops and wearables, while the larger ones suit on-premises servers and embedded enterprise services[1]. PFN showed the 0.8B model translating on a smartphone in airplane mode[1].

Users can feed the model surrounding context and a glossary for names, technical terms and product names[1]. For Japanese output it can be told to use formal honorific forms or plain language[1]. Because nothing leaves the device or server, PFN pitches it for contracts, R&D documents and support logs that can't go to an outside API[1].

PFN says even the 0.8B model outperforms comparable translation LLMs such as Google's TranslateGemma and Tencent's Hy-MT2, across Japanese-English, Japanese-to-multilingual and English-to-multilingual tests[1]. Those comparisons are PFN's own and include proprietary benchmarks, with no independent replication yet[1].

The models use the PLaMo Community License[2]. Personal and non-commercial use, including academic work, is free with no revenue limit and no registration[2]. Commercial use is also free when the user and its affiliates had revenue of 1 billion yen (about US$6.5 million) or less in the most recent fiscal year, but it requires submitting PFN's commercial-use registration form first[2][1]. Bigger companies need a separate licence agreement with PFN[2]. That makes it open weights rather than open source in the strict sense.

Translation is one of the few AI jobs that people need most when they're offline: travelling, in a disaster zone, or handling documents that legally can't leave the building. A credible 0.8B model that runs on a phone, from a lab with deep Japanese-language expertise, gives app builders a real alternative to sending every sentence to a cloud API.