PLaMo 3 Translate: Offline Translation You Can Actually Ship — Explained Honestly

Preferred Networks (PFN) of Japan released PLaMo 3 Translate today, 5 November 2026: a family of translation-specialized large language models with open weights, in four sizes from 0.8B to 31B parameters, available now on Hugging Face. The headline is simple enough to fit in one sentence: high-quality translation that runs entirely on your own device, with no internet connection at all.

I have not tested these models hands-on, and this article does not pretend otherwise. Everything below comes from PFN's own announcement, checked today. What I can do is the part that matters more than the press release: work out what "offline translation" actually means for a developer building a real product, which size to pick, and where the licence catches you.

What PLaMo 3 Translate actually is

PLaMo 3 Translate is not a general-purpose chatbot that happens to translate. It is a translation-specialized model, built on PLaMo 3 NICT — a pre-trained model developed through joint research between PFN and Japan's National Institute of Information and Communications Technology (NICT). It is the same engine that powers PFN's own PLaMo Translate service, now released as downloadable weights you can run yourself.

Four sizes are available today on Hugging Face, under the pfn-plamo repositories:

All four cover 53 languages, including English and Japanese. And all four run fully offline — on smartphones, laptops, and on-premises servers — with no internet needed. PFN's demo makes the point in the most direct way possible: the smallest 0.8B model translating on a phone in airplane mode, at normal speed.

Why offline is the whole story

Most translation today means sending your text to somebody else's cloud. That is fine for a tourist asking for directions. It is a problem the moment the text is a contract, an R&D document, or a customer support log — the exact examples PFN names. With PLaMo 3 Translate, neither the original text nor the translation ever leaves the device or server you control. Nothing is sent to an external cloud server, because there is no cloud server in the loop at all.

That changes which products are even possible. A travel app that translates on a plane. A support tool that translates tickets inside a company's own network. A field device for disaster response that works where the network does not. PFN names tourism, enterprise business, and disaster response as the use cases it expects — and for once the list reads like a real description of where offline matters, not a marketing slide.

There is a second, quieter benefit: latency and cost stop depending on an API. No per-character bill, no round trip, no rate limit, no outage on somebody else's status page. The cost becomes the hardware you already have.

The claim to check: the 0.8B beating bigger names

PFN's boldest claim is about the smallest model. According to PFN's own comparative evaluations, even the 0.8B version outperforms comparable translation-specific LLMs such as Google's TranslateGemma and Tencent's Hy-MT2.

Read that sentence carefully, because the attribution matters. This is PFN's evaluation, including proprietary benchmarks — not an independent test, and not a public leaderboard result. The chart in PFN's announcement plots model size against evaluation score for Japanese–English, Japanese-to-multilingual, and English-to-multilingual translation, and the PLaMo models sit in the "smaller yet higher quality" corner. That is encouraging. It is also exactly the kind of claim that needs independent hands-on testing before anyone builds a product on it.

What I can say honestly: if the claim holds up even approximately, the 0.8B is the interesting one. A sub-billion-parameter model that translates well on a phone changes the economics of on-device apps in a way a 31B model never will. The 31B exists for the teams that need the best quality PFN can offer and have the servers to run it. The 0.8B is the one that could end up inside millions of apps.

The features that matter for real apps

Translation quality is only half the job in a real product. The other half is control, and PFN built three controls in that developers will recognise as the difference between a demo and something shippable:

• Glossary control. You can feed the model a glossary for proper nouns, technical jargon, and product names, so your brand and your terminology translate the same way every time. Anyone who has watched a product name get "translated" into nonsense knows why this matters.
• Context-aware output. You can give the model surrounding context, so the same sentence translates differently in a support ticket and in a marketing page — because it should.
• Japanese tone control. When translating into Japanese, you can specify formal honorific forms or plain language forms. Japanese is a language where getting the politeness level wrong is not a small error; it is the whole message. This is a feature that only exists because the model was built by a Japanese team for real Japanese use.

None of these are flashy. All of them are the kind of thing you discover you need in week two of building, when the demo is already done.

The licence catch — read this before you ship

Here is the part of this release that most coverage will bury, so I am putting it plainly.

PLaMo 3 Translate is not released under a fully open licence like Apache 2.0 or MIT. It uses the PLaMo Community License. Under it, individuals and organisations with annual revenue of 1 billion yen or less (about $6.5M USD) in their most recent fiscal year can download and use the models free of charge, for commercial and non-commercial use alike. That covers internal corporate operations and integration into commercial products and services.

Organisations above that revenue threshold need a commercial licence agreement with PFN.

So: if you are an indie developer, a startup, or a small company, this is effectively free for your product. If you are a large company, it is not free — talk to PFN first. PFN even issued a correction to its own announcement on the day of release to make the "1 billion yen or less" wording exact, which tells you how seriously they take the boundary.

Do not ship a product on these weights without checking which side of that line you are on. The weights are open; the licence is not unconditional.

Which size should you use?

• 0.8B — start here. If you are building a mobile or on-device feature, this is the model to test first. PFN's claim is that it punches above its weight; your test on your own language pairs and your own content is the one that counts. Run it on a real phone, on real user text, and compare it against whatever you use today.
• 2B — the middle step. For laptops and light server use where the 0.8B is close but not quite good enough, and you still want something small and fast.
• 8B — the quality step. For server-side translation where quality matters more than model size, and you have the hardware for it.
• 31B — the top of the family. For on-prem enterprise deployments where you want the best translation PFN offers and you have the servers to run a 31-billion-parameter model.

PFN's own framing is that you pick by your requirements for translation quality, latency, processing volume, and target hardware. That is the right way to think about it: the four sizes are not good-better-best in the abstract; they are four different answers to "what hardware do I have, and what quality do I need?"

The limits — stated plainly

• No independent verification yet. Every quality claim in this article is PFN's claim, from PFN's evaluation. Nobody outside PFN has published hands-on results at the time of writing. The TranslateGemma and Hy-MT2 comparisons come from PFN's proprietary benchmarks.
• Translation only. This is a translation-specialized model, not a general-purpose LLM. Do not expect it to write your essays or answer general questions.
• The licence has a revenue ceiling. Free up to ¥1B annual revenue; above that, a commercial licence is required. This is not Apache/MIT, and treating it as if it were would be a mistake.
• 53 languages, not every language. English and Japanese are covered, along with 51 others — but if your product needs a language outside the list, this is not your model.
• No hands-on testing was performed for this article. I have not run any of the four sizes. The airplane-mode demo is PFN's, and the quality chart is PFN's.

The bottom line is a practical one. PLaMo 3 Translate is a genuinely useful release, and its most useful part is not the model size or the benchmark chart. It is that a translation model you can actually ship — offline, on a phone, with glossary control and privacy built in — now exists as downloadable weights, free for small teams and indie developers. Try the 0.8B on your own text first, on a real device, in airplane mode. If it translates your content well there, you have just removed a cloud dependency from your product. If it does not, you have learned that in an afternoon, for free.

Have you tried running a translation model fully offline before — and what broke first, the quality or the device? That is the test I would trust more than any launch chart, and I would like to hear how it went.

Sources: Preferred Networks — "PFN Releases Open-Weight Translation LLM PLaMo 3 Translate, Bringing World-Class Translation AI to Edge Devices", 5 November 2026, https://www.preferred.jp/en/news/pr20261105. Models on Hugging Face: pfn-plamo/plamo-3-translate-0.8b, -2b, -8b, -31b. All facts checked 5 November 2026. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.