What Preferred Networks released
Tokyo-based Preferred Networks (PFN) published the weights for PLaMo 3 Translate on 5 November 2026. The release has four sizes: 0.8B, 2B, 8B and 31B parameters. All four are translation-specialised language models, and all are on Hugging Face under the pfnet organisation. According to Impress Watch's report, they're the same models behind PFN's commercial PLaMo Translate web service. PFN is pitching them for offline use on phones, PCs and in-house servers, so source text never has to leave the machine.
The models translate between 53 languages. The 31B model card lists them, and Simplified and Traditional Chinese are counted separately. Japanese and English are the core pair. The list also includes several languages widely spoken in Auckland: both forms of Chinese, Korean, Hindi, Tagalog, Vietnamese, Punjabi and Indonesian. Te reo Māori, Samoan and Tongan aren't on it, and we found no statement from PFN about plans to add Pacific languages.
How the models were built
PFN started from PLaMo 3 NICT, a base model it pretrained jointly with Japan's National Institute of Information and Communications Technology, and then specialised it for translation. The PFN engineering post describes a training set of about 37.3 billion tokens. PFN mined source text from the web and generated the translations with Gemma 4 31B IT, using three techniques. Tournament-style synthesis has the model fuse candidate translations in pairs. Back-translation produces forward and reverse pairs. Iterative self-distillation feeds the model's improved outputs back into training.
Training ran in two supervised fine-tuning (SFT) phases: a large first phase of about 33.9 billion tokens and a smaller, curated second phase of about 3.4 billion tokens that includes browser and file translation tasks. Two rounds of direct preference optimisation (DPO) followed. DPO is a method that trains on pairs of preferred and rejected outputs without a separate reward model. The 8B and smaller models were then distilled from the 31B teacher. PFN also reports trying GRPO, a reinforcement learning method. GRPO raised benchmark scores, but human reviewers rated the output less fluent, so PFN dropped it. A lab rarely publishes that kind of trade-off, and it's a useful warning for anyone tuning a translation model against automatic metrics.
What the benchmarks do and don't show
PFN claims its 0.8B model beats translation-specialised models more than 30 times its size on Japanese-English and Japanese-to-other-language pairs, and that the 2B model does the same on English-to-other-language pairs. The comparison models are Google DeepMind's TranslateGemma and Tencent's Hy-MT2. Impress Watch repeats the claim and notes that the article itself gives no benchmark details.
Builders should read the results with care. Two of PFN's four evaluation sets (PFTB and a SaaS set that covers only English to Japanese) are internal. The public sets are WMT24++, scored with COMET-22, and FLORES+, scored with chrF++. PFN says it gave WMT24++ less weight during development because the set is mostly short sentences and has some reference quality problems. Gemma 4 31B IT generated the synthetic training data and also ranked the DPO pairs, and PFN measured the DPO gains with Gemini 3.1 Pro as judge. Those gains were modest: 53 to 59 percent win rates per round. We haven't seen an independent evaluation yet. For a language pair you care about, run your own test set before relying on the model.
Formats, sizes and hardware
Each size ships in several formats. FP8, W4A16, W4AFP8 and NVFP4 builds target vLLM on different generations of NVIDIA GPU. GGUF builds run under llama.cpp on CPUs or GPUs, and MLX 4-bit builds run on Apple silicon via mlx-lm. The 2B model card lists weight files of about 5.76 GB in BF16 and 1.81 GB as GGUF Q4_K_M, which suits a modest laptop. At the top end, the 31B card lists 63.51 GB in BF16, 19.53 GB as GGUF Q4_K_M and 18.66 GB as MLX 4-bit. At 4-bit, the 31B model is within reach of a 32 GB Apple silicon Mac or a 24 GB GPU, but PFN doesn't publish RAM or VRAM requirements, and context length adds memory on top of the weights.
The 8B GGUF card publishes how much accuracy each quantisation level costs compared with BF16. Q4_K_M (5.28 GB) stays within a few thousandths on PFN's in-house scores and moves FLORES+ chrF++ by less than 0.1. Q2_K (3.45 GB) drops chrF++ by 0.91 on the multilingual pairs. PFN recommends starting with Q4_K_M.
Every size was trained with a combined input and output limit of 32,768 tokens. The model card advises splitting longer documents, and translation history counts toward the same limit.
Running it locally
The 8B GGUF card gives these commands to download the model and serve it through llama.cpp's OpenAI-compatible server:
pip install -U huggingface_hub
hf download pfnet/plamo-3-translate-8b-GGUF plamo-3-translate-8b-Q4_K_M.gguf --local-dir plamo-3-translate-8b-GGUF
cd plamo-3-translate-8b-GGUF
llama-server -m plamo-3-translate-8b-Q4_K_M.gguf -c 16384 --port 8000 --alias plamo-3-translate --jinja --no-prefill-assistantThe -c 16384 flag sets the combined context and can be lowered to fit the memory you have. The card says recent GPU-enabled llama.cpp builds offload layers to the GPU automatically, and -ngl sets the layer count by hand. It asks for a recent build but doesn't name a minimum version. For NVIDIA servers, the model cards show vllm serve with --trust-remote-code, --max-model-len 16384 and --max-num-seqs 4.
The models use PFN's own PFML3 prompt format, which has special tokens, and the dataset field must always be set to translation. Glossaries are lists of source and target term pairs. In Transformers, glossaries and free-text context are passed as arguments to apply_chat_template(). Over vLLM, the 31B card shows only language options going through chat_template_kwargs. To send a glossary or context to vLLM, you write the PFML3 prompt yourself and call the Completions API. Code that assumes a generic chat API will need a small adapter.
Licence terms
The weights are released under the PLaMo Community License, which isn't an OSI open-source licence. Personal, academic and other non-commercial use is free with no registration. Commercial use is free for organisations whose most recent annual revenue, including affiliates, was 1 billion yen or less, but they must register with PFN in advance. Organisations above that threshold need a separate commercial agreement. Most New Zealand startups and small firms will fall under the threshold. Larger companies, and any business whose parent group is large, should check the revenue figure before building a product on the models.
What builders here could do with it
The most direct use is translating confidential material without a cloud API. Examples include contracts and technical documents for Japanese trading partners, or customer messages in Chinese, Korean or Hindi, translated on a server you control. Glossary support helps keep product names, legal terms and te reo Māori place names unchanged in the output. The 0.8B and 2B builds are small enough to trial inside a desktop app, or to run on a CPU-only edge box for a kiosk or field device that works offline.
Japanese-English translation has the strongest evidence behind it. For other pairs, especially smaller languages like Cebuano or Khmer, build an evaluation set from your own documents and compare against your current provider before switching. If you build something on PLaMo 3 Translate, the society's projects page is open for members to share it.
Sources
- PLaMo 3 Translate: World-Class Translation with Open Weights (Preferred Networks)
- pfnet/plamo-3-translate-2b model card (Hugging Face)
- pfnet/plamo-3-translate-8b-GGUF model card (Hugging Face)
- pfnet/plamo-3-translate-31b model card (Hugging Face)
- Impress Watch report on PLaMo 3 Translate, 5 November 2026
- ThursdAI: Open Source and Open Weights
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.