# PLaMo 3 Translate Opens Local Translation Weights, With Commercial License Conditions

Preferred Networks released four translation model sizes for 53 languages. Local runtimes, benchmark coverage and commercial terms matter more than the demo.

By Paras Madan | 2026-11-05


A translation feature has to fit somewhere: a browser, a laptop, a company server or a cloud service. The best model on a chart can still be the wrong choice if it cannot run where the text needs to stay.

Preferred Networks' [November 5 release of PLaMo 3 Translate](https://www.preferred.jp/en/news/pr20261105) gives developers four open-weight options: 0.8B, 2B, 8B and 31B parameters, supporting 53 languages. The models specialize in translation and build on the PLaMo 3 NICT base developed with Japan's National Institute of Information and Communications Technology.

## What developers can use

The [model card](https://huggingface.co/pfnet/plamo-3-translate-0.8b) lists GGUF variants for llama.cpp and MLX variants for Apple silicon alongside GPU-oriented formats. It supports glossary entries, surrounding context and plain or polite Japanese output. The documented training limit is 32,768 tokens across input and output.

These are useful ingredients for a controlled evaluation. A glossary can help test whether product names and specialist terms survive a translation. Context lets a team check consistency across paragraphs rather than judging isolated sentences. Neither feature removes the need to examine omissions, altered numbers or changes in meaning.

PFN's [technical report](https://www.preferred.jp/en/blog/tech/plamo-3-translate-weight-release) includes a smartphone demonstration running in airplane mode. It explicitly says this is an internal demo and that the company has not released that app. Developers can experiment with the weights; readers should not expect a ready-made phone application from the demonstration.

For a local deployment, test the complete application with network access disabled. That checks the behavior of the wrapper, dependencies and update process as well as inference. Hosting a model locally creates the possibility of keeping text on the device, but the surrounding software still determines where data travels.

## Read the benchmark coverage carefully

PFN's technical evaluation combines internal Japanese-English tests with WMT24++ and FLORES+. Its multilingual comparison covers a selected 15-language set through Japanese and English, rather than every possible pair among all 53 supported languages. The reported comparisons are vendor evaluations.

A practical test set should therefore follow the intended product. Evaluate the actual language directions, document lengths and terminology users need. Ask qualified reviewers to mark meaning-changing errors separately from awkward phrasing. Include difficult cases such as names, dates, negation and instructions that must remain unambiguous.

Measure latency and memory on the intended hardware, too. A smaller model that meets the quality requirement may be easier to deploy widely. A larger one may justify its cost for passages that need more careful review. The choice should come from those measurements, not parameter count alone.

## Commercial use needs a license check

The model card describes a custom PLaMo Community License. Non-commercial use, including academic use, is free without registration. Commercial use requires advance registration for the free tier. The card applies its ¥1 billion-or-less threshold to the latest-fiscal-year revenue or sales of the user or its affiliates. Above that threshold, PFN requires a separate agreement.

The [full community agreement](https://plamo.preferredai.jp/info/plamo-community-license-en) also sets attribution requirements for redistribution and treats models trained on its outputs as derived models subject to the same license conditions. Teams planning to generate training data should review that provision before starting a large job.

Those conditions belong in an adoption decision early. A working prototype does not settle whether a company, its affiliates or a redistributed product fits the permitted use. Review the linked terms before committing to a commercial integration.

PLaMo 3 Translate is a concrete addition to the options for local translation. Its value will depend on a combination that broad model rankings rarely capture: the right languages, acceptable errors, practical hardware and suitable rights.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn model releases and deployment tradeoffs into clear stories developers can use.

[Book a 30-minute call](https://cal.com/paras-madan/30min).