PFN Releases PLaMo 3 Translate Weights: Above ¥1 Billion, Start with the License

On 5 November, Preferred Networks (PFN) announced that it had published the weights of its translation model, PLaMo 3 Translate, on Hugging Face. The announcement lists four sizes: 0.8B, 2B, 8B, and 31B. The model supports 53 languages, can take surrounding context and a glossary, and can be set to plain or polite Japanese. PFN’s account is that, when the model runs on a device or server the user controls, the source text and the translation need not be sent to an external server.

The prior question is who may use it for business. The PLaMo Community License defines a commercial purpose as any use beyond personal or academic use. The definition includes use inside the user’s organization, and developing, modifying, or providing products, services, or business operations. Commercial use has to meet two conditions together: prior registration on PFN’s registration page, and annual revenue or sales of the user or an affiliated company that do not exceed 1 billion yen in the most recent fiscal year. Above that line, the license requires a separate commercial license from PFN. The English technical blog says noncommercial use, including academic use, has no income or revenue limit and does not require prior registration.

The “1 billion yen or less” in the news is checked against the license

The body of the Japanese announcement says that companies and individuals whose income or sales were 1 billion yen or less in the most recent fiscal year may, under the community license, use the model free of charge for commercial and noncommercial purposes. A correction posted the same day at 11:45 writes “under or less than 1 billion yen,” which does not match the body or the license phrase “must not exceed.” Read with the English blog, free commercial use requires income or sales of 1 billion yen or less in the most recent fiscal year, plus prior registration. Above that line, a separate agreement is required.

The line that can be checked clause by clause is Article 3 of the license: commercial use requires prior registration, and the annual revenue or sales of the user or an affiliated company must not exceed 1 billion yen. Use inside an organization is already inside the definition of a commercial purpose. Downloading the weights does not complete those two conditions.

Offline is a way of deploying the model, and the license still applies

The announcement describes contracts, research documents, and customer correspondence as work that can stay local: machine translation under the company’s own information rules, without sending the source text to an external server. The technical blog also offers BF16, FP8, 4-bit, GGUF, and MLX formats for different devices and inference engines.

Those passages describe where the model runs. The license still applies to a model running locally. The airplane-mode phone demonstration in the announcement uses the smallest 0.8B model and is labeled as a demonstration environment at normal speed. It shows that a small model can produce a translation on a device without a network. It does not report a measured result for a company’s document workflow.

The quality comparison is PFN’s own evaluation

The technical blog places the finished models on one chart with Google DeepMind’s TranslateGemma and Tencent’s Hy-MT2. Japanese–English scores are COMET-22 averaged across PFTB, SaaS, and WMT24++. Multilingual scores are FLORES+ chrF++ averages. The blog says the 0.8B model surpasses the translation-specialized models on that chart for Japanese–English and for Japanese with other languages, including models with more than 30 times as many parameters. For English with other languages, the 2B model already surpasses those comparison models. A footnote says the evaluation covers a subset of the supported languages.

The same blog presents “the world’s best translation-specialized LLM” as the authors’ judgment from these results. Article 7 of the license says the model and its output are provided as is, and that PFN gives no warranty of accuracy or fitness for a particular purpose. Output must not be the sole basis for a legal matter, a medical diagnosis, a financial decision, or an important personal assessment. The translation model is built by PFN on PLaMo 3 NICT, a pretrained model from joint research by PFN and Japan’s National Institute of Information and Communications Technology.

Read the attribution and liability clauses before handing work to a third party

The license groups the model, derived models, and output together as model materials. Article 4 says that providing those materials to a third party requires a copy of the license and the words “Built with PLaMo” on the related website, interface, blog, product information, or documentation. Article 5 says that public distribution of an output must state that it is the output of the model or a derived model. A model trained on that output is a derived model under the same license.

Article 8 says that when the user applies the model materials to business, PFN assumes no liability in connection with the agreement and the model materials. For use that is not business use, the liability cap is 500 yen, except for intent or gross negligence. Before a model is built into a customer-facing translation feature, or a translation is handed to a client, those two articles sit alongside the registration and revenue conditions.

Four gates before connecting the model to internal translation

• License gate: Use inside an organization is a commercial purpose under the license. Free commercial use requires prior registration, and the income or sales of the user or an affiliated company must not exceed 1 billion yen in the most recent fiscal year. Above that line, obtain a separate commercial license first.
• Data gate: Source text stays inside only when the model actually runs on a device or server the company controls. Publishing the weights on Hugging Face does not by itself keep a later call on the premises.
• Quality gate: The comparison with TranslateGemma and Hy-MT2 is PFN’s evaluation, and it covers only some of the supported languages. A contract, a medical text, or a financial text still needs a check beyond the model output.
• Delivery gate: Before providing output to a third party, check the attribution requirements in Articles 4 and 5, and the business-use liability terms in Article 8.

PLaMo 3 Translate puts a locally runnable translation model in the hands of anyone who downloads it, and it keeps commercial use inside a registration step and a revenue line. A company taking work abroad, and considering the model for internal documents or client delivery, should first check affiliate income or sales, whether registration is complete, and which notices the license asks for when a translation goes to a third party. Ongoing translation work still needs terminology, a quality target, and a named owner of the final text, agreed inside the professional translation workflow.

Sources: the Preferred Networks announcement of 5 November 2026; the PFN technical blog, “PLaMo 3 Translate: World-Class Translation with Open Weights”; the English text of the PLaMo Community License. The quality scores come from PFN’s own evaluation. The license gives no warranty of accuracy for the model or its output.