Japanese AI Firm Preferred Networks Releases Open-Weight Translation Model PLaMo 3 Translate for Offline Mobile Use
Preferred Networks has released PLaMo 3 Translate, an open-weight LLM series supporting 53 languages, with the smallest 0.8B model capable of running offline on smartphones, claiming superior performance over existing competitors.
Japanese artificial intelligence company Preferred Networks (PFN) released its open-weight translation large language model, "PLaMo 3 Translate," on the Hugging Face platform on November 5, 2026. The model family comes in four parameter sizes—0.8B, 2B, 8B, and 31B—and is designed to enable localized translation on end-user devices such as smartphones and laptops without requiring an internet connection.
PLaMo 3 Translate supports mutual translation between 53 languages, including Japanese and English. According to benchmark results published by PFN, even the smallest 0.8B model outperforms existing specialized translation models like Google's TranslateGemma and Tencent's Hy-MT2 in tasks involving Japanese. If these claims hold up, it signifies a shift towards bringing high-quality multilingual translation capabilities to resource-constrained edge devices.
The models are released under the "PLaMo Community License." Non-commercial use is free. For commercial use, companies with annual income or revenue below JPY 1 billion (approx. CNY 47 million) can use the models for free after registering with PFN. Commercial use above this threshold requires a separate license agreement.
Technically, PFN stated that training data was refined through techniques such as tournament-based data synthesis, back-translation, and iterative self-distillation, followed by two stages of supervised fine-tuning (SFT) and direct preference optimization (DPO). Models of 8B parameters and smaller also utilized knowledge distillation from the 31B model to enhance their performance. Various quantization formats, including BF16, FP8, and W4A16, are provided to suit different hardware environments.