Is PLaMo 3 Translate state of the art?Pending benchmarks.Announced Nov 4, 2026; no independent results yet.
Preferred Networks' open-weight translation-specialized LLM family (0.8B, 2B, 8B, 31B), built on PLaMo 3 NICT from joint work with Japan's NICT and further trained on tens of billions of translation tokens. Supports 53 languages; optional glossary and surrounding-context input; Japanese output can request plain or polite form. Weights on Hugging Face under the PLaMo Community License, with GGUF, MLX, FP8 and other edge formats so the smaller sizes can run locally on phones and PCs. No separate public inference API for these weights is stated; PFN's hosted PLaMo Translate service is a different product.
The details so far
- Lab
- Preferred Networks
- Announced
- Nov 4, 2026
- API
- Not yet
- Open weights
- Published
- Parameters
- 0.8B / 2B / 8B / 31B (family)
- Role
- Translation-specialized causal LLMs: source text in, target-language translation out, with optional glossary, context and translation history
- Sizes
- 0.8B, 2B, 8B and 31B; each size ships BF16 plus quantized/converted variants (FP8, W4A16, W4AFP8, NVFP4, GGUF, MLX-4bit)
- Languages
- 53 languages including Japanese and English (Simplified and Traditional Chinese counted separately; HF language chip shows 52 because it groups Chinese)
- Context window
- 32,768 tokens for input and output combined during training; lab recommends splitting longer passages
- Prompt format
- PFML3 (dataset=translation) via chat template or handwritten prompts; Transformers, vLLM, llama.cpp and mlx-lm paths documented on the cards
- Base
- Continued from PLaMo 3 NICT (PFN + NICT joint pretrained model), then translation-focused training
- Checkpoints
- pfnet/plamo-3-translate-0.8b, -2b, -8b, -31b on Hugging Face (collection pfnet/plamo-3-translate)
- License
- PLaMo Community License (HF license tag: other / license_name plamo-community-license). Personal and non-commercial use free with no registration; commercial use free when the user's most recent fiscal-year revenue or sales is ¥1B or less after registration; above that threshold requires a separate commercial agreement with PFN
- Related
- Separate from older PLaMo foundation models, the hosted PLaMo Translate service, and PLaMo 3 NICT (the pretrained base)
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: on PFN's Japanese–English suite (average of PFTB, SaaS and WMT24++ COMET-22) and FLORES+ multilingual chrF++ averages, even the 2B model sits above the translation-specialized LLMs PFN compared against on every metric in the published figure (PFN tech blog / model cards).
Lab claim: PFN comparative evaluations place the 0.8B model above Google TranslateGemma and Tencent Hy-MT2 among translation-specific LLMs of similar size (press release).
Lab claim: BF16 31B scores include PFTB (En↔Ja) 0.6244, SaaS (En→Ja) 0.5786, WMT24++ COMET-22 (En→Ja) 0.8799, FLORES+ chrF++ Ja↔Multilingual 37.81 and En↔Multilingual 56.61 (31B model card).
Sources
LABPFN press: PLaMo 3 Translate open-weight releasepreferred.jp/en/news/1820 · Nov 5, 2026LABPLaMo 3 Translate tech blog (Preferred Networks)preferred.jp/en/blog/tech/plamo-3-translate-weight-release · Nov 5, 2026REPOpfnet/plamo-3-translate collection on Hugging Facehuggingface.co/collections/pfnet/plamo-3-translateREPOpfnet/plamo-3-translate-31b model card (Hugging Face)huggingface.co/pfnet/plamo-3-translate-31bLABPreferred Networks on X: PLaMo 3 Translatex.com/PreferredNet/status/2118136522700640295 · Nov 4, 2026LABPreferred Networks Japan on X: PLaMo 3 Translatex.com/PreferredNetJP/status/2118132607116517500 · Nov 4, 2026Tracking continues until an independent board lists it.