pfnet open-sources PLaMo 3 Translate translation models in four sizes

According to a post seen on the Reddit community LocalLLaMA on November 7, 2026, at 13:43, pfnet has open-sourced a translation-specialized LLM called PLaMo 3 Translate, publishing its weights on Hugging Face. The model family is offered in four sizes: 0.8B, 2B, 8B, and 31B. The official description claims that even the smallest 0.8B variant outperforms existing translation-specialized LLMs, including models with more than 30 times as many parameters, on Japanese-to-English and Japanese-to-other-languages translation. The developers attribute this result to meticulous data refinement, highlighting techniques such as tournament-based data synthesis and back-translation as the key factors behind the performance. That is where the available record stops: the single source available covers the release and the published claims, and provides no information about accompanying technical reports, licensing terms, or follow-up updates, so the timeline can only be drawn up to the point where the weights went public on Hugging Face.