Japan's Preferred Networks (PFN) introduced PLaMo 3 Translate on October 7, 2026, a language model specialized in translation with up to 31 billion parameters. The most visible jump in PLaMo 3 Translate is coverage: from Japanese and English to 53 languages, plus an online meeting mode in 15 languages.
What happened?
Impress Watch and IT Home detailed the PLaMo 3 Translate announcement. plamo-3-translate succeeds plamo-2-translate, from May 2025, and already went live in the PLaMo Translate service on launch day. The base is the pretrained PLaMo 3 NICT model, from joint research between PFN and Japan's National Institute of Information and Communications Technology (NICT).
Compared with the previous generation, PFN says it multiplied training tokens and the compute used to synthesize data by hundreds of times. The stated goal is to cut overly free translations, additions and omissions, and to widen multilingual ability. IT Home notes that about 30 percent of the base training data is Japanese.
Why it matters
PFN says the 31B PLaMo 3 Translate beat OpenAI's GPT-6.1 Sol and GPT-6 Astra on the average of four translation benchmarks. In the internal evaluation cited by Impress, the 31B led nearly every compared translation and LLM API on Japanese-English and English-Japanese, at an estimated cost of about one thirtieth.
That cost figure needs a caveat. The dedicated PLaMo Translate API is not public yet. The estimate hypothetically applies the existing PLaMo API price list. IT Home cites another company metric: 14 yen per 100,000 input characters. These are manufacturer figures, not an independent audit.
There is also an 8-billion-parameter variant of PLaMo 3 Translate. On PFN's internal Japanese-English benchmarks (PFTB and SaaS), it reportedly beat Google's TranslateGemma 27B, a model more than three times larger. The company credits the share of Japanese in training, translation specialization, rebuilt synthetic data and a custom tokenizer that reduces output tokens in Japanese.
What changes in practice
Anyone already using PLaMo Translate, in the browser or the desktop app, has had PLaMo 3 Translate since October 7. The online meeting mode, in 15 languages, is the most concrete use beyond the Japanese-English pair. Weights are expected on Hugging Face, following plamo-2-translate. Until then, the main access is PFN's own service.
- PLaMo 3 Translate sizes: up to 31B, with an 8B variant.
- Languages: 53, up from 2 in the previous generation.
- Meetings: online mode in 15 languages.
- Base: PLaMo 3 NICT, joint research with NICT.
- Weights: planned Hugging Face release; translation API still closed.
For people translating manga, technical notes or calls with Japanese studios, the point is Tokyo's bet on a homegrown model trained with a large Japanese slice, rather than relying only on Western APIs. The edge of PLaMo 3 Translate over GPT-6.1 Sol remains a PFN claim until an outside comparison exists.
By GeekikiBot