PFN's largest disclosed pretraining run, with NICT: a 120B-parameter base (weight-reuse initialization from PLaMo 3 31B, 3T tokens, ~10^24.3 FLOPs) scoring pfgen-bench 0.83 / JMMLU 0.77 — and a distilled, publicly released 31B (1.5T tokens, 256K-context continued pretraining). The 120B weights are not public; the open 31B carries its distilled capability. A meaningful scale-up for Japan's sovereign-model efforts beyond the API-only PLaMo 3.0 Prime line.

PLaMo 3 Translate (open weights 2026-10-27) is a translation-specialized family built on PLaMo 3 NICT and further trained on "tens of billions of tokens of translation-focused data". It comes in 0.8B, 2B, 8B and 31B sizes, each also shipped in FP8, W4A16, W4AFP8, NVFP4, GGUF and MLX formats. It covers 53 languages, accepts surrounding context and a glossary, and was trained at up to 32,768 input-plus-output tokens. PFN reports that even the 2B beats TranslateGemma and HY-MT2 on every metric it shows. The 31B scores 0.8799 WMT24++ COMET-22 for English to Japanese, and 4-bit formats stay within about 0.01 of BF16. The license is the PLaMo Community License, free for non-commercial use.

Model Details

Parameters 120B
Training tokens 3T

Variants

Name Parameters Notes
plamo-3-nict-2604-120b-base 120B weights not public
plamo-3-nict-2604-31b-base 31B distilled, open weights, 256K-context CPT
plamo-3-translate (0.8B / 2B / 8B / 31B) 31B translation-specialized, 53 languages, PLaMo Community License (2026-10-27)
japaneseopen-weightsovereign

Related