PLaMo 3 Translate: Japan's open translation model runs offline, even on an iPhone

What's going on

Preferred Networks (PFN), one of Japan's best-known AI companies, released PLaMo 3 Translate as an open-weight model on November 5, 2026. The language model is built specifically for translation. You can download the weights from Hugging Face in four sizes: 0.8B, 2B, 8B and 31B (0.8 to 31 billion parameters). PFN calls its translation quality "world-class". The same model also powers PFN's paid "PLaMo Translation" service, which moved to the third model generation on October 7, 2026.

What sets it apart: the model does translation and nothing else. It's not built for chatting. The small versions are compact enough to run on smartphones and laptops without an internet connection, so your text never has to go to a cloud service.

What the model can do

PLaMo 3 Translate supports 53 languages. PFN's multilingual tests include German, English, French, Spanish, Italian, Dutch, Russian, Chinese (Simplified and Traditional), Korean, Arabic, Thai, Vietnamese and Indonesian. Japanese and English are clearly the focus, though.

It also has some practical options:

• Surrounding context: You can pass in the text around the passage you want translated, so the model gets references right.
• Glossary: You can specify how proper names and technical terms should be translated.
• Japanese style: You can pick the polite "desu/masu" form or the neutral "da/de aru" form.
• In PFN's own web page translation test, HTML tags stay intact. That matters if you translate whole pages with a browser extension, for example.

PFN says the model cards on Hugging Face explain how to use these options.

The four sizes and a hands-on test

Japanese tech outlet Gigazine tried out the small variants right away. The 0.8B model ran on an iPhone 17 Pro in the PocketPal AI app with Q4_K_M quantization. Quantization shrinks a model by storing its numbers at lower precision. It translated the opening of Jane Austen's "Pride and Prejudice" (about 280 words) into Japanese in 0.515 seconds, and the first chapter of Einstein's book on relativity in 0.387 seconds. Gigazine described the output as very natural Japanese with no broken grammar, and called the offline speed "pretty shocking". The team also tested the 2B model in LM Studio on a laptop without a dedicated GPU, using the abstract of the famous Transformer paper "Attention Is All You Need". It was also fast and highly accurate.

PFN itself shows an internal demo of the 0.8B model translating on a smartphone in airplane mode. According to its tech blog, though, PFN doesn't currently offer a dedicated app.

How PFN trained the model

In its tech blog, PFN explains that the key wasn't new training methods so much as better data. The starting point is a third-generation PLaMo base model. According to an explainer on note.com, it's "PLaMo 3 NICT", developed together with Japan's NICT research institute, with about 30 percent Japanese training data.

Here's how the training pairs (source text plus translation) were created:

• Collecting source texts: PFN filtered large web datasets for diverse text written by native speakers. That meant removing spam pages, already machine-translated text and mixed-language text. There are even dedicated filters, for example to keep Malay out of Indonesian text or to catch Traditional Chinese text that was auto-converted from Simplified Chinese.
• Tournament data synthesis: Google's open model Gemma 4 31B IT produced several translations per text. These went head to head in pairs. The winner advanced but borrowed good phrasings from the loser. Round by round, the translation got better.
• Back-translation: The model also learns the reverse direction, from the AI translation back to the human-written original. That teaches it to write fluent, natural sentences instead of just imitating translationese.
• Iterative self-distillation: Interim versions of PLaMo 3 Translate produced new translations themselves. Whenever these beat the existing training data, they replaced it, and the model was retrained.

All told, this added up to about 37.3 billion tokens (chunks of text): 33.9 billion for base training, followed by 3.4 billion especially high-quality tokens from academic papers, news, technical documents and web pages. Next came DPO (Direct Preference Optimization), where the model learns from pairs of "better vs. worse translation". For the 31B model, this step ran in two stages. Gigazine gives an example: after this fine-tuning, "seeing happy customers made it all worth it" came out as noticeably more natural Japanese. The models up to 8B also learned from the 31B model through distillation, with the big model acting as teacher for the small ones.

PFN measured performance with two in-house benchmarks (PFTB for long, demanding Japanese–English texts; SaaS for web pages with HTML) plus the public WMT24++ and FLORES+ tests. PFN says it prefers its own benchmarks because WMT24++ consists of short sentences with some questionable reference translations.

Formats and license

Every size comes in several formats. PFN used AutoRound, among other tools, for quantization and says the quality loss compared with the uncompressed BF16 version is minimal across all formats.

The model is released under the PLaMo Community License. Personal, hobby, learning and research use is free, with no registration required. Companies and individuals can use it commercially for free after registering through a form, as long as their revenue in the last fiscal year was no more than 1 billion yen (ITmedia says "under" 1 billion yen). Larger companies need a separate license agreement. So this isn't a classic open-source license.

What you need to try it

• 0.8B: runs on current smartphones (tested: iPhone 17 Pro with PocketPal AI, Q4_K_M) and on practically any PC or Mac, even CPU-only. The 4-bit file is under 1 GB.
• 2B: according to Gigazine, runs smoothly in LM Studio on a laptop without a GPU; about 2–3 GB of RAM is enough.
• 8B: best with a GPU with roughly 6–8 GB of VRAM or more (4-bit); on a Mac, 16 GB of unified memory or more using the MLX version. It works on a CPU, just slower.
• 31B: in 4-bit, roughly 20 GB of VRAM (e.g. RTX 3090/4090) or a Mac with 32 GB or more. The FP8 and NVFP4 variants need newer Nvidia cards.

You can also test the small and mid-sized variants on Google Colab's free GPU.

Our take

PLaMo 3 Translate is one of the strongest cases yet for local translation. A 0.8B model that produces natural Japanese on a phone in a fraction of a second makes offline translation realistic for travel, disaster response and confidential company documents. Those are exactly the use cases PFN itself names. Its direct competitors are Google's TranslateGemma and Tencent's Hy-MT2, both also openly available.

There are still caveats. All the performance comparisons come from PFN itself, mostly on in-house benchmarks; independent measurements are still missing. The focus is clearly on Japanese. The sources don't show how pairs like German–French or German–English stack up against DeepL or large chat models. Gigazine's positive hands-on tests covered English–Japanese. And the license is generous but comes with a revenue cap and mandatory registration for commercial use. Still, if you work a lot with Japanese or want a fast, private translation model for low-end hardware, it's worth a try.