PFN opens PLaMo 3 Translate weights, 0.8B to 31B

PFN's release covers the full range of deployment targets rather than a single flagship. The 0.8B model is aimed at phones and other limited hardware, the 2B at small computing environments, the 8B at GPU-equipped PCs and internal servers, and the 31B at the highest translation quality in the series. Quantized variants in GGUF, MLX and FP8 formats accompany the BF16 versions, so users can match the model to their hardware.

The quality gains are presented as a data problem as much as a modeling one. PFN built training pairs by having Google's Gemma 4 31B IT translate web text, then used "tournament-style data synthesis" to compare translations two at a time and keep the better one. It also trained models to reverse the direction (generating the original language from the translation) and used iterative self-distillation, where the model's own improved outputs replace training data. The final dataset totaled about 373億 tokens, split into about 339億 for base translation ability and about 34億 of high-quality data from academic papers, news, technical documents and web pages.

PFN also compared the models with existing translation-focused LLMs including Google's TranslateGemma and Tencent's Hy-MT2. In PFN's benchmark, even the 0.8B model exceeded models with more than 30× the parameters on Japanese-English and Japanese-to-multilingual translation. The license terms are another factor for business users: free non-commercial use, free commercial use with registration for companies at or below 10億円 in annual sales, and a separate contract above that threshold.