Contents

Preferred Networks (PFN) released the model weights for PLaMo 3 Translate on November 5, 2026. Available in four sizes from 0.8B to 31B parameters, the models support translation across 53 languages. Organizations can run them on their own hardware and keep translation processing in-house, but should review the license before deploying them for commercial use.

Key Points#

  • PLaMo 3 Translate comes in four sizes—0.8B, 2B, 8B, and 31B—with model weights available on Hugging Face.
  • It supports translation across 53 languages, with options to specify surrounding context, glossaries, and Japanese writing style.
  • PFN reports benchmark performance exceeding TranslateGemma and Hy-MT2, though these comparisons come from the developer.
  • Noncommercial use is free and requires no registration. Commercial terms depend on annual income or revenue, and a separate agreement may be required.

Running Locally#

PLaMo 3 Translate is a translation-focused language model that can be deployed on devices or company-managed servers. If inference runs entirely within that environment, source texts and translations need not be sent to external translation servers.

This makes it suitable for work involving contracts, research documents, and customer communications subject to restrictions on external transmission. PFN also envisions integration into products and services. The models already power its PLaMo translation service, and the release of the weights lets users operate them independently.

As an example of use on small devices, PFN shared an internal demo of the 0.8B model translating on a smartphone in airplane mode. This demonstrates offline use but does not guarantee compatibility or practical speeds on every smartphone.

Choosing a Size and Format#

The four sizes let you balance model capabilities against deployment resources. Start with the smallest model that meets your translation requirements, then compare it with larger models using the same documents before upgrading your hardware.

Alongside unquantized BF16 weights, PFN provides models quantized in formats including FP8, W4A16, W4AFP8, and NVFP4. GGUF and MLX-4bit versions are also available. Quantization can reduce memory requirements, but the appropriate download format depends on your inference software and hardware.

You can evaluate the models as follows:

1. Open PFN’s PLaMo 3 Translate collection on Hugging Face. 2. Choose a model size and a format supported by your runtime environment. 3. Follow the model card to check how to run it and configure translation options. 4. Measure translation quality, memory usage, and speed using representative documents.

Controlling Terminology and Style#

In addition to Japanese–English translation, the models support translation across 53 languages, including Chinese, Korean, and French. You can supply surrounding context, which helps translate passages whose meaning depends on nearby text.

You can provide a glossary to specify translations for product names, proper nouns, and technical terms. For Japanese output, you can choose plain or polite style. These features help with translations that must follow an organization’s terminology and editorial rules.

Test these controls explicitly. Include passages with recurring technical terms and passages whose meaning depends on the preceding sentence. Check that the output preserves meaning, applies the glossary consistently, and maintains the requested style throughout.

Reviewing Scores and Licensing Carefully#

According to PFN, the 0.8B model outperformed all comparison models, including TranslateGemma 27B and Hy-MT2 30B-A3B, in Japanese–English and Japanese–multilingual evaluations. PFN also reports leading results for models of 2B and above in English–multilingual translation. The Japanese–English results average three benchmarks, while the multilingual results use FLORES+. These scores provide a reason to try the smaller models, but do not guarantee quality for every language pair or specialized document.

The model weights are covered by the PLaMo Community License. According to PFN, noncommercial use—including personal learning, hobbies, and academic use—is free and requires no advance registration. Publicly available weights should therefore not be interpreted as permission for unrestricted commercial use.

For commercial use, users or their affiliates that fall within the specified annual income or revenue threshold of ¥1 billion can use the models for free by registering in advance through PFN’s designated form. Those exceeding this threshold require a separate license agreement. Before deployment, review the full license and confirm how its definitions apply to your organization.

Next Steps#

Run the 0.8B model on a small set of documents approved for use that contain your actual terminology. Record translation errors, processing time, and memory usage before evaluating larger sizes.

Source: https://gihyo.jp/article/2026/11/plamo-3-translate-open-weights

Contents