On November 5, PLaMo 3 Translate was released as an open-weight model, with Japan's Preferred Networks (PFN) handing developers the translation-specialized large language model used in its own translation service, published on Hugging Face. What was released is not an online service, but a set of model weights that can be installed directly on devices, covering 53 languages including Japanese and English. PFN says the aim is to make high-quality translation possible on edge devices, with use cases including tourism, business, and disaster response.
| Size | Parameter count | | 0.8B | 800 million parameters | | 2B | 2 billion parameters | | 8B | 8 billion parameters | | 31B | 31 billion parameters |
How to choose among the four sizes
The four sizes are 0.8B, 2B, 8B, and 31B, and developers choose based on translation quality, speed, volume, and device. PFN demonstrated the smallest 0.8B model running offline on a phone in airplane mode, meaning it can produce translations with no network connection. In PFN's own comparative evaluation, even the smallest model scored above Google's TranslateGemma and Tencent's Hy-MT2, two other translation-specialized models, though this is the publisher's own benchmark and independent reproduction is still to be seen. The model also supports fixing translations of proper nouns and technical terms through a glossary, and adjusting output using context; when translating into Japanese, users can specify polite or plain style, a practical detail for business documents and internal papers.
Where the free-use boundary lies
The license terms are the line in this news that most needs a careful read. Under the PLaMo Community License, companies and individuals whose revenue or sales in the most recent fiscal year were 1 billion yen or less can download and use the model free of charge, for commercial and non-commercial purposes alike, whether on a phone, a laptop, or a company server. Commercial use beyond that scope requires a separate commercial license from PFN. The model itself was developed from PLaMo 3 NICT, a pretrained model obtained through joint research between PFN and Japan's National Institute of Information and Communications Technology (NICT). Another practical benefit is confidentiality: because the model runs on a server or device the user manages, source and translated texts do not need to be sent to an external server, so restricted documents such as contracts, research materials, and customer correspondence can be translated entirely in-house.
