PLaMo 3 Translate

What it is

PLaMo 3 Translate is a family of language models built specifically for translation, from the Japanese company Preferred Networks (PFN). It comes in four sizes — 0.8B, 2B, 8B and 31B parameters — and translates among 53 languages.^[1] The weights of the largest, 31B version were uploaded to Hugging Face on 26 October 2026.^[1]

Parameter count is a rough measure of how much capacity a model has. The 0.8B model is small enough to run on modest hardware, while the 31B one belongs to the class of models served on GPUs. Because all four sizes come from the same training recipe and ship in the same set of formats, the smaller ones are not just demonstrations: PFN says even the 2B model beats the translation-specialised models it compared against on every metric it measured.^[1]

Where it comes from

The starting point is PLaMo 3 NICT, a pretrained model developed jointly by PFN and Japan's National Institute of Information and Communications Technology.^[1] PFN then trained that base further on tens of billions of tokens of translation data.^[1] In other words, the translation ability is layered onto a general pretrained model rather than assembled from scratch for the translation task alone. That is the usual pattern for specialised models today: take a broad base that already carries a lot of language knowledge, and concentrate the extra training on the behaviour you want.

Beyond one sentence at a time

The model can take surrounding context and a glossary into account.^[1] For practical translation work this matters in two ways. Context lets it keep referents and pronouns consistent across a passage instead of translating each sentence in isolation; a glossary fixes how a particular term must be rendered, which is the usual requirement when an organisation has an approved terminology list. For Japanese output the model offers plain or polite forms.^[1] Training used a maximum of 32,768 tokens for input and output combined,^[1] so the source passage, any context and glossary supplied with it, and the translation itself all have to fit inside that budget.

Benchmark results

The 31B model, measured in BF16 precision, scores 0.6244 on PFTB, 0.5786 on SaaS, and 0.8799 COMET-22 on WMT24++.^[1] COMET-22 is a learned metric: instead of counting matching words, it compares a candidate translation against reference translations and rewards preserved meaning, with higher scores better. On FLORES+ the model reaches 37.81 chrF++ for Japanese against 14 other languages and 56.61 for English.^[1] chrF++ is a character-level overlap score, which suits languages such as Japanese that do not put spaces between words. The two figures cover different sets of language pairs and are not directly comparable with each other.

Formats and licence

Every size ships in FP8, W4A16, W4AFP8, NVFP4, GGUF and MLX variants.^[1] Reduced-precision formats store weights in fewer bits, which lowers the memory a model occupies and usually speeds up inference, at some cost in numerical accuracy; GGUF and MLX are the formats used by common local runtimes, while the others target serving on GPUs.

The licence is the PLaMo Community License: free for personal and academic use, with separate terms for commercial use.^[1][2]

Where it sits

Among the 53 languages, the reported evaluation keeps Japanese and English apart rather than giving one aggregate number, and the FLORES+ table is organised around Japanese against 14 other languages.^[1] Together with the plain and polite Japanese output option, that places the model's centre of gravity in Japanese-language work. The release pattern — four sizes, each in six formats, under a licence that separates personal and academic use from commercial use — is aimed at letting the same family be tried on a laptop and deployed on a server without changing model family.