Preferred Networks Releases PLaMo 3 Translate as Open Weights for Offline Translation in 53 Languages
Preferred Networks (PFN) on November 5 published the model weights of PLaMo 3 Translate, the translation-specialised large language model behind its PLaMo Translate service, on Hugging Face, the Tokyo company announced. The model handles 53 languages including Japanese and English and comes in four sizes of 800 million, 2 billion, 8 billion and 31 billion parameters, so users can trade quality against speed and hardware. PFN says the smallest version runs on a smartphone in airplane mode.
Under the PLaMo Community License, individuals and companies whose revenue in the latest fiscal year is 1 billion yen or less can download the models and use them for commercial or non-commercial purposes on phones, laptops or in-house servers without going through the cloud; larger users need a commercial licence from PFN. Mado no Mori reported that companies using the free commercial route must pre-register through a dedicated form, and that PFN also ships quantised builds plus GGUF files for llama.cpp and MLX files for Apple silicon.
PFN says its own evaluations, which include internal benchmarks, show even the smallest model beating Google’s TranslateGemma and Tencent’s Hy-MT2. Users can pass surrounding context and glossaries for product names and technical terms, and can choose between polite and plain styles when translating into Japanese. The base model is PLaMo 3 NICT, built in joint research with the National Institute of Information and Communications Technology (NICT).
What it means in practice
The release extends PFN’s on-premises push with PLaMo 3.1 Prime and fits a wider domestic effort to test homegrown models, such as the generative-AI evaluation platform that Sakura Internet, PFN and NICT are rolling out. The practical draw is confidentiality: contracts, research papers and customer correspondence can be translated without sending text to outside servers, which matters for manufacturers and local governments wary of cloud tools. PFN names tourism, business and disaster response as early use cases.
What to watch is how independent tests compare with PFN’s self-reported scores, whether device makers and app developers embed the small models in their products, and how many companies above the revenue threshold sign paid licences.