The Pulse
PFN Releases Open-Weight PLaMo 3 Translate Models in Four Sizes
Preferred Networks has released four open-weight translation models, ranging from 0.8 billion to 31 billion parameters, for use on devices and servers without an internet connection. The release describes offline translation across 53 langu
AI.info Team · Written with AI · fact-checked ·
A 0.8B model for offline translation
Preferred Networks (PFN) announced the open-weight release of PLaMo 3 Translate on November 5, 2026. The suite contains four translation-specialized large language models, developed originally for PFN’s PLaMo Translate service. Their weights are available on PFN’s Hugging Face page.
The models have 0.8 billion, 2 billion, 8 billion and 31 billion parameters. PFN says developers can choose among them based on translation quality, latency, processing volume and the hardware they plan to use. The company says applications incorporating the models can translate across 53 languages, including English and Japanese, without an internet connection.
PFN demonstrated the smallest model translating on a smartphone in airplane mode, running at normal speed in a demo environment. The release describes the smaller models as suitable for smartphones, laptops and wearable devices, and says use can scale to on-premises enterprise servers and embedded corporate services. The models provide weights that users can run locally on their own hardware.
License terms and availability
All four models are available on Hugging Face under the PLaMo Community License. The release’s general license statement says individuals and enterprises with annual revenue of 1 billion Japanese yen or less in their most recent fiscal year can download and use the models free of charge for commercial and non-commercial applications.
The commercial-licensing section describes a different threshold: it says commercial use is permitted free of charge for organizations generating less than 1 billion Japanese yen annually, and that organizations exceeding the threshold need a commercial license agreement with PFN. The release therefore gives different wording for the revenue limit in its general license statement and its commercial-use section. It says organizations can test the models’ speed and quality on their own devices before integrating them into workflows.
Evaluation, privacy and customization
PFN says its comparative evaluations found that the smallest, 0.8B model outperformed comparable translation-specific LLMs, including Google’s TranslateGemma and Tencent’s Hy-MT2. The company says its evaluations include proprietary benchmarks. Its comparison covers Japanese-English translation, Japanese-to-multilingual translation and English-to-multilingual translation.
PFN says local execution keeps both original text and translated output from being sent to external cloud servers. The company cites contracts, research and development documentation, and customer support logs as examples of sensitive material that could be translated locally while maintaining corporate privacy and security policies.
Users can provide surrounding context and glossaries for proper nouns, technical terminology and product names, according to the release. For translations into Japanese, PFN says users can choose formal honorific forms or plain language.
Based on PLaMo 3 NICT
PFN says PLaMo 3 Translate is based on PLaMo 3 NICT, a pre-trained model developed through joint research between PFN and Japan’s National Institute of Information and Communications Technology. The company says it aims for the release to expand recognition of its PLaMo foundation model and support open innovation through edge-device LLMs. PFN expects the release to drive use cases in tourism, enterprise business, disaster response and other sectors.