PLaMo 3 Translate

PLaMo 3 Translate is a translation-specialized LLM from Japan's Preferred Networks (PFN) and the model behind its commercial PLaMo Translation service. On 2026-11-05 PFN published open weights for all four sizes on Hugging Face (pfnet): 0.8B, 2B, 8B and 31B.

It is built on PLaMo 3 NICT, a base model co-developed by PFN and Japan's NICT, further trained on tens of billions of translation tokens. It translates among 53 languages, accepts surrounding context and glossaries, and lets you choose plain or polite style for Japanese output. PFN says even the 0.8B model beats translation-specialized LLMs more than 30× its size on Japanese–English and Japanese–other-language pairs.

What problem it solves

Confidential documents (contracts, medical records, internal material) often can't be sent to a cloud translation API. PLaMo 3 Translate lets companies and individuals translate locally on phones, laptops and on-prem servers, even offline.

Sizes and formats

Size	Typical use
0.8B / 2B	On-device translation on phones and light laptops
8B	High-quality local translation on one GPU or a Mac
31B	Highest-quality server-side translation

Each size ships in BF16, FP8, W4A16, W4AFP8, NVFP4, GGUF and MLX-4bit, runnable with Transformers, vLLM, llama.cpp or mlx-lm. Max training length was 32,768 tokens for input plus output.

How it was built

Tournament-based data synthesis, back-translation and iterative self-distillation, combined with SFT, DPO and distillation into the smaller sizes.

License and limits

• PLaMo Community License: free for personal and non-commercial (including academic) use; companies with ≤ JPY 1 billion revenue in the last fiscal year can use it commercially for free after registering; larger companies need a separate license.
• Translation-only; not meant as a general chat model.
• Split very long texts so input plus output stays within the training length.

Who it's for

• Companies that need high-quality Japanese↔English/Chinese translation without data leaving their network;
• Mobile developers embedding offline translation in apps;
• Localization, subtitling and document-translation professionals.

Getting started

1. Read PFN's tech blog and pick a size.
2. Mac: mlx_lm.generate --model pfnet/plamo-3-translate-8b-MLX-4bit; PC: load pfnet/plamo-3-translate-8b-GGUF in llama.cpp.
3. Server: vllm serve pfnet/plamo-3-translate-31b-FP8.
4. Follow the model card's prompt format for source/target language, context and glossary; check registration or licensing before commercial use.