PFN opens the weights of its translation-focused LLM PLaMo 3 Translate, letting the 0.8B model translate 53 languages on an offline phone
Synopsis
PFN released PLaMo 3 Translate, the translation-focused large language model behind its PLaMo translation service, as open-weight models in four sizes, 0.8B, 2B, 8B and 31B, which can be embedded in phones, laptops or in-house servers to translate among 53 languages including Japanese and English without a network connection. By PFN's own evaluation, the smallest model scores above Google's TranslateGemma and Tencent's Hy-MT2 on translation quality, and companies and individuals with revenue or sales of 1 billion yen or less in the latest fiscal year may use it commercially for free under the PLaMo Community License.
Interpretation
A translation-focused large model no longer has to be reached only through a cloud service: PFN published the weights of PLaMo 3 Translate on Hugging Face so users can run the model locally on their own phones, laptops or in-house servers. Previously this translation capability was offered as PFN's PLaMo translation service, used through the service; the same model family is now downloadable as open weights in four sizes, 0.8B, 2B, 8B and 31B, chosen by quality, speed, throughput and device. PFN demonstrated translation running on a phone in airplane mode with no network connection using the smallest model, PLaMo 3 Translate 0.8B, and states that this smallest model scored above Google's TranslateGemma and Tencent's Hy-MT2 in its own evaluation.
The models are tiered by device compute: the small sizes target phones, laptops and wearable devices, while the large sizes target in-house servers or product integration, so one model family covers configurations from personal devices to enterprise deployment. Size selection maps to four needs, translation quality, response speed, throughput and the device in use, rather than offering a single general-purpose model. The four published model identifiers are plamo-3-translate-0.8b, plamo-3-translate-2b, plamo-3-translate-8b and plamo-3-translate-31b, all released on Hugging Face under the PLaMo Community License.
Local execution keeps source and translated text from leaving the organization's own devices or servers, so contracts, R&D materials and customer correspondence that should not be sent out can still be translated under the organization's own information-management rules. Cloud translation requires sending text to an external server, whereas local deployment keeps translation processing inside the organization. PFN states that when the model runs on self-managed servers or devices, translation processing can be completed in-house without sending source or translated text to external servers.
Perspective
The demonstrated scope is the release of four open-weight model sizes on Hugging Face under the PLaMo Community License and a translation demo of the smallest 0.8B model on a phone in airplane mode with no network connection; commercial use beyond the community license requires a commercial license agreement with PFN.
The translation-quality comparison is based on PFN's own evaluation including an in-house benchmark, and the material gives no specific scores, test-set composition or per-size results; the demo ran at normal speed in a demonstration environment.