Preferred Networks has released PLaMo 3 Translate weights for running translation on users’ own devices and servers. The models support 53 languages and build on joint research with Japan’s information and communications technology institute. Context and glossary controls help specify wording, while a community licence sets separate conditions for non-commercial and commercial use.
Preferred Networks, a Japanese AI developer, released PLaMo 3 Translate weights through the model-sharing platform Hugging Face on 5 November. The models specialize in translation across 53 languages, including Japanese and English. They draw on technology already used in the company’s online translation service. The release gives users access to the underlying weights for running that translation technology on hardware they control.[1], [2]
The underlying PLaMo 3 NICT model comes from joint research with Japan’s National Institute of Information and Communications Technology, followed by additional translation training. Smaller versions target phones and laptops, while larger models suit internal servers. A company demonstration uses a phone in airplane mode. With local deployment, source text and translated output can be processed without sending them to an external server.[1]
Users can provide surrounding text as context and set preferred translations for names and technical terms through a glossary. Japanese output offers formal and plain styles. The PLaMo Community License permits free non-commercial use. Commercial use within its revenue threshold requires prior registration; larger organizations are directed to a separate agreement. The commercial-use threshold is annual revenue of 1 billion Japanese yen or less in the latest fiscal year.[1]
