PLaMo 3 Translate Review: Free Offline AI Translation in 53 Languages

Your phone can now translate 53 languages with no signal, no subscription, and no data leaving the device. On November 5, 2026, Preferred Networks (PFN) of Tokyo released the weights for PLaMo 3 Translate, a translation-specialized language model built for exactly that job. It comes in four sizes, from a 0.8-billion-parameter model small enough to run on a phone in airplane mode to a 31-billion-parameter version for servers. The price is zero for most users: individuals and businesses with annual revenue of ¥1 billion or less can download and use it free, including inside commercial products. The catch sits in the license fine print, and it matters if you plan to ship it in a product. This review covers what it does, what it costs, how good PFN says it is, and where the claims still need independent proof.

What PLaMo 3 Translate is

PLaMo 3 Translate is a suite of open-weight models that do one thing: translate text between languages. PFN built it on PLaMo 3 NICT, a base model developed through joint research with Japan's National Institute of Information and Communications Technology (NICT). The same model family already powers PFN's PLaMo Translate service, which Japan's Digital Agency adopted for its government AI project. What changed on November 5 is that the weights are now public: anyone can download them from Hugging Face and run them on their own hardware, with no API key and no connection to PFN's servers.

That last point is the whole pitch. Cloud translation services send your text to someone else's servers. PLaMo 3 Translate runs where you run it, so contracts, medical records, and support logs never leave the device. PFN names tourism, enterprise business, and disaster response as the first use cases, and the last one is easy to see: translation that works when the network is down.

The models cover 53 languages, including English and Japanese. They accept surrounding context with each request, take glossaries for proper nouns and technical terms, and let users pick formal honorific or plain-language forms when translating into Japanese. That tone control matters: Japanese business writing lives or dies on politeness, and most machine translation tools give you one register and no choice.

The four sizes, and which one to download

PFN ships four models. The smaller ones trade some quality for speed and hardware that fits in a pocket; the larger ones trade convenience for accuracy. Each size also comes in several formats, so the same model can run on a GPU server, a laptop CPU, or a Mac.

The format matters as much as the size. BF16 is the full-precision version for GPUs running Transformers or vLLM. GGUF runs through llama.cpp on CPUs and GPUs, which is the practical route for a laptop. MLX-4bit targets Apple silicon through mlx-lm. Quantized variants (FP8, W4A16, W4AFP8, NVFP4) shrink the models for specific NVIDIA hardware, and PFN says its benchmarks show they keep nearly the same accuracy as the full BF16 weights.

The headline demo is the 0.8B model translating in real time on a smartphone in airplane mode. PFN is careful about what that demo is: an internal demonstration, not an app you can download. There is no PLaMo Translate app in the app stores. To use the model on a phone today, a developer has to build the app around the weights. That is the honest state of things on launch day.

The license: free, with one catch

PLaMo 3 Translate is released under the PLaMo Community License, and the price is zero for most people. The terms split three ways:

• Personal use: free, no revenue limit, no registration. Download the weights and use them for study, hobbies, or personal projects.
• Non-commercial use: free, no revenue limit, no registration. Academic use falls here.
• Commercial use: free if you or your affiliates had revenue of ¥1 billion or less in the most recent fiscal year, provided you register in advance through PFN's dedicated form. Above that threshold, you need a separate commercial license agreement with PFN.

PFN puts the ¥1 billion threshold at roughly $6.5 million. At November 5, 2026 market rates of about ¥158 to the dollar, it works out to about $6.3 million, or roughly £4.8 million and C$8.8 million. Those are approximations, and the yen figure is the one that counts: ¥1 billion or less in your most recent fiscal year. PFN corrected its announcement on launch day, changing "less than ¥1 billion" to "¥1 billion or less", so a business sitting exactly at the line is inside the free tier.

Two details deserve attention. First, the advance registration for commercial use is a real step: the tech blog states commercial use below the threshold is free "provided that the user registers in advance through the dedicated form." Second, this is an open-weight release under PFN's own community license, not an open-source release under a standard OSI license, so read the terms on the PLaMo Community License page before building a product on it.

There are no per-character fees, no API meter, and no subscription. Once the weights are on your hardware, translating a million characters costs the same as translating ten: the electricity to run your own machine. That is where the money argument for PLaMo 3 Translate is strongest.

How good is it? PFN's numbers, with the caveats attached

PFN's central claim is striking: even the smallest 0.8B model outperforms comparable translation-specialized models, including models with more than 30 times as many parameters, in Japanese-to-English translation and translation between Japanese and other languages. The press release names Google's TranslateGemma and Tencent's Hy-MT2 as the models it beats. If it holds up independently, a phone-sized model beating 25-billion-parameter rivals is a real step forward.

Every part of that claim needs the same qualifier: these are PFN's own evaluations. The company tested on four benchmarks: two public ones, WMT24++ and FLORES+, and two of its own, the PFTB benchmark for Japanese-English translation across domains like academic papers and poetry, and a SaaS benchmark built to mirror real requests to its translation service. PFN says it leaned mainly on its own two benchmarks during development because it had concerns about reference quality in WMT24++. That is transparent, and it means the strongest numbers come from tests PFN designed itself. No independent evaluation had been published at the time of writing, so treat the "beats models 30 times larger" line as a claim to watch, not a settled result.

What is documented is how the models were built. Training started from the PLaMo 3 NICT base model, then ran through two stages of supervised fine-tuning, first on a large volume of translation pairs, then on a smaller, higher-quality set, followed by direct preference optimization, a final step that teaches the model to prefer better translations over worse ones. The 8B model and smaller ones were distilled from the finished 31B model, which is how the small models inherited much of the big model's quality. The training data was largely synthesized: PFN collected source texts from the web, then generated translations with a large language model, using back-translation and tournament-style data selection to clean the results.

One more honest limit: the model is text-only. It does not transcribe speech or translate spoken conversations by itself. Speech-to-text is a different job, and PFN's release covers written translation only.

How to get it running

All four models are live on Hugging Face, with usage instructions on each model card:

• plamo-3-translate-0.8b on Hugging Face
• plamo-3-translate-2b on Hugging Face
• plamo-3-translate-8b on Hugging Face
• plamo-3-translate-31b on Hugging Face

For a first try, the GGUF version of the 0.8B model with llama.cpp is the simplest path on a laptop, and the MLX-4bit version is the equivalent on a Mac. Server users will want the BF16 or FP8 weights with vLLM. Commercial users under the ¥1 billion threshold should register through PFN's form before shipping anything, and anyone above the threshold should contact PFN about a commercial license rather than assuming the free tier applies.

If you only need occasional translation and you are happy sending text to the cloud, this is not the tool for you. The open weights pay off when privacy, offline use, or volume makes the cloud awkward: a law firm translating contracts, a hospital handling patient records, or an app translating thousands of product listings a day. For those users, the download is the product.

PLaMo 3 Translate vs cloud translation APIs

The fair comparison is against the paid APIs businesses actually use, where the bill grows with every character translated. The rates below are 2026 list prices reported in API pricing roundups.

The table flatters PLaMo on price and privacy, because those are the dimensions where an offline open-weight model wins outright. The cloud services win on the other two columns. Google covers 249 languages against PLaMo's 53, and all three work through a simple API call with no hardware to manage and no registration form. A team translating 500,000 characters a month pays nothing to Google or DeepL either, thanks to their free tiers, so PLaMo's price advantage only bites at volume, or where privacy rules out the cloud entirely. The honest way to read it: PLaMo 3 Translate is the cheapest option at scale and the only option that keeps data fully in-house, but it asks you to do the hosting work yourself.

There is also a nearby comparison on the open-model side. Google's TranslateGemma is a free open translation model in the same spirit, and it is one of the two models PFN names as a benchmark rival. PFN claims its 0.8B model beats TranslateGemma while running on a phone, which, if independent tests confirm it, makes PLaMo the stronger pick for on-device use. Until then, both are free and open-weight, and PLaMo is the one with the phone demo and the Japanese-tone controls.

Common mistakes to avoid

Assuming "free" means "no paperwork." Personal and academic use needs no registration. Commercial use under the ¥1 billion threshold does. Register first, ship second.

Downloading the 31B model for a laptop. The big model is for servers with serious GPU memory. Start with the 0.8B or 2B GGUF builds unless you know your hardware can carry the larger weights.

Expecting an app. PFN demonstrated the 0.8B model on a phone but has not released a phone app. Right now this is weights and model cards, not a consumer product. Developers can build on it; everyone else waits for someone to ship an app around it.

Treating PFN's benchmarks as settled fact. The "beats 30x larger models" claim comes from PFN's own tests, weighted toward its own benchmarks. It is a reason to try the model, not a reason to quote it as proven.

Using it for speech. PLaMo 3 Translate handles written text. Spoken translation needs a speech model in front of it, a separate piece of work. Our review of NVIDIA's Nemotron 3.5 Transcribe covers the transcription half of that pipeline.

Who should use it

The clearest fit is a small business or developer team that translates a lot, cares about privacy, and sits under the ¥1 billion revenue line. A company translating product documentation, a clinic handling patient forms, or a developer building a travel app that works in airplane mode all get a capable model for the cost of hosting it. Larger companies above the revenue threshold can still use it, but they need to talk to PFN first, and at that point the price is a negotiation, not a download.

For everyone else, the release still matters. It shows that a model small enough for a phone can now claim parity with far larger translation models, at least on its maker's tests. Open-weight releases like this one, alongside downloadable models such as Meta's Muse Spark Open and Kimi K3.1 Pro's open weights, are steadily moving capable AI from rented APIs onto hardware people already own. If independent tests back PFN's numbers, offline translation stops being a compromise and starts being the default.

Is PLaMo 3 Translate really free?

Yes, within the PLaMo Community License. Personal and non-commercial use is free with no revenue limit and no registration. Commercial use is free if your revenue was ¥1 billion or less, about $6.5 million, in the most recent fiscal year, provided you register in advance. Above that threshold you need a separate commercial license.

Does PLaMo 3 Translate work without the internet?

Yes. The weights run entirely on your own device, and PFN demonstrated the 0.8B model translating on a smartphone in airplane mode. You need internet once to download the model; after that, it works offline.

Which size should I download?

Start with the 0.8B model for on-device or laptop use, and move to 2B or 8B for higher quality if you have the hardware. The 31B model is for GPU servers. GGUF with llama.cpp is the simplest route on a laptop, and MLX-4bit is the equivalent on Apple silicon.

How many languages does it support?

53 languages, including English and Japanese, fewer than Google Cloud Translation at 249, so check your language pair is covered before committing.

Can I use it in a commercial product?

Yes, if your revenue was ¥1 billion or less in the most recent fiscal year and you register in advance through PFN's form. Above that threshold, contact PFN for a commercial license agreement.

Is there a PLaMo Translate app for my phone?

Not yet. PFN's phone demo was an internal demonstration, and the company says it has not released an app. The November 5 release is model weights on Hugging Face, so using it on a phone today requires building an app around those weights, or waiting for someone else to.

Sources

• Preferred Networks, "PFN Releases Open-Weight Translation LLM PLaMo 3 Translate, Bringing World-Class Translation AI to Edge Devices," November 5, 2026 — preferred.jp/en/news/pr20261105
• Preferred Networks, "PLaMo 3 Translate: World-Class Translation with Open Weights, Here's How We Built It," November 5, 2026 — preferred.jp/en/blog/tech/plamo-3-translate-weight-release
• PFN, PLaMo 3 Translate 0.8B model card — huggingface.co/pfnet/plamo-3-translate-0.8b
• PFN, PLaMo 3 Translate 31B model card — huggingface.co/pfnet/plamo-3-translate-31b
• PLaMo Community License — plamo.preferredai.jp
• AIB, "PFN Releases PLaMo 3 Translate, a Translation-Specialized LLM," November 5, 2026 — aib.vote/en/news/pfn-plamo-3-translate-weights-release
• ChatsControl, "Translation API Pricing 2026: DeepL vs Google Cloud vs Azure" — chatscontrol.com/blog/translation-api-pricing-2026-deepl-google-azure
• MigraPay, USD to JPY mid-market rate, November 5, 2026 — migrapay.com/exchange-rate/usd-to-jpy