On Nov. 5, 2026, Preferred Networks (PFN) issued a Tokyo-dated press release titled “PFN Releases Open-Weight Translation LLM PLaMo 3 Translate, Bringing World-Class Translation AI to Edge Devices.” PFN describes an open-weight release of translation-specialized large language models originally built for its PLaMo Translate service, with weights published on PFN’s Hugging Face hub. For related coverage, see IndiaAI Releases EKA-1 and Offline-Capable Dhīr AI Models and Tencent Releases Hy4 With Vision and Computer Use.
Key facts
| Item | Detail |
|---|---|
| Announced | Nov. 5, 2026 (Tokyo) |
| By | Preferred Networks, Inc. (PFN) |
| Models | PLaMo 3 Translate: 0.8B, 2B, 8B, 31B parameters |
| Languages | 53 languages including English and Japanese (when integrated into device apps) |
| License | PLaMo Community License; commercial license above revenue threshold |
| Weights | Public on Hugging Face (pfnet) |
| Foundation | Built on PLaMo 3 NICT (PFN–NICT joint research) |
What PFN released
PFN says PLaMo 3 Translate is a suite of translation-specialized LLMs whose weights are now publicly available after development for PFN’s PLaMo Translate service. The release provides model weights so teams can run translation on their own hardware rather than relying on a hosted service alone.
Four sizes are listed: 0.8B (0.8 billion parameters), 2B, 8B, and 31B. PFN positions the tiered lineup so developers can trade off translation quality, latency, processing volume, and target hardware.
Offline edge use and languages
By integrating PLaMo 3 Translate into device applications, PFN says users can achieve high-quality, high-speed translations across 53 languages including English and Japanese without needing an internet connection. The release also states that users can run the models entirely on edge devices such as smartphones, laptops, and internal corporate servers without requiring an internet connection.
PFN’s materials include a demonstration of the 0.8B model performing translation on a smartphone in airplane mode.
Quality claims and benchmarks
According to PFN’s comparative evaluations, even the smallest 0.8B version of PLaMo 3 Translate outperforms comparable translation-specific LLMs, such as Google’s TranslateGemma and Tencent’s Hy-MT2. PFN ties that claim to its own evaluations, including proprietary benchmarks, and illustrates model size versus evaluation score in charts for Japanese-English, Japanese-to-multilingual, and English-to-multilingual settings.
Those comparisons reflect PFN’s testing setup; this article does not extend them beyond what the release reports.
Features PFN highlights
PFN lists several themes for the suite. It cites high-quality output on small models in its comparative evaluations. It notes scalability across hardware ecosystems, from smartphones and wearables offline to on-premises servers. It emphasizes privacy when models run locally, stating that neither source text nor translated output is sent to external cloud servers in that setup.
PFN also describes advanced glossary control and contextual formatting, including feeding surrounding context and glossaries for proper nouns and technical terms. For Japanese output, PFN says the model can target formal honorific forms or plain language forms.
Licensing and downloads
All four models are available on Hugging Face under the PLaMo Community License, PFN says. Under that license, individuals and enterprises with an 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 both commercial and non-commercial applications, per the Nov. 5 release (PFN posted a correction at 11:45 on Nov. 5 clarifying the revenue threshold wording).
Organizations exceeding that revenue threshold require a commercial license agreement with PFN, the release states. Commercial use at or below the threshold—including internal corporate operations and integration into commercial products or services—is free once users submit PFN’s PLaMo Commercial Use Registration Form, per the Hugging Face model card; non-commercial and academic use is free with no revenue limit or registration.
Model repositories named in the release:
Development background
PFN notes that PLaMo 3 Translate is a specialized translation model built on PLaMo 3 NICT, a pre-trained model from joint research between PFN and Japan’s National Institute of Information and Communications Technology (NICT). PFN says it expects the open-weight release to expand recognition of the broader PLaMo foundation model and to accelerate open innovation through edge-device LLMs, with possible use cases in tourism, enterprise business, disaster response sectors, and more.
What it means
Analysis (ours). PFN is packaging translation as downloadable weights sized for phones through on-premises servers, with a revenue-capped community license that still allows commercial use below the threshold. The headline quality claim rests on PFN’s own benchmarks against named translation-focused models, not independent leaderboard scores. For teams that must keep text on-device, the offline and privacy framing is the practical hook; the 31B tier is the option when accuracy beats battery life.
What to watch
Independent replication of PFN’s comparative charts, commercial-license uptake above the 1 billion yen line, and integrations in travel, enterprise, or emergency workflows PFN names as examples.