PLaMo 3 Translate
PFN's translation model in 0.8B to 31B sizes for 53 languages, with a 0.57 GB GGUF and a revenue-limited community license.
In short. PLaMo 3 Translate is a family of AI translation models from the Japanese company Preferred Networks. You pick a size, from a 0.8B model small enough for a laptop CPU to a 31B model for a GPU server. They translate between 53 languages, can follow a glossary and surrounding context, and can write Japanese in a plain or a polite style. The license is free for personal and academic use, but a company has to register first and needs a separate deal above 1 billion yen of revenue.
Summary
Preferred Networks (PFN) published the weights of PLaMo 3 Translate on Hugging Face and described them in a blog post dated 5 November 2026. There are four sizes, called 0.8B, 2B, 8B and 31B, which hold about 0.89, 2.88, 8.53 and 31.76 billion parameters. The models start from PLaMo 3 NICT, a pretrained model that PFN developed jointly with Japan's National Institute of Information and Communications Technology, and PFN trained them further on tens of billions of tokens of translation data. They cover 53 languages, with Simplified and Traditional Chinese counted separately, and the target language is required while the source language is detected if you leave it out. The card supports surrounding context, a glossary and a choice between plain and polite Japanese. Training used 32,768 tokens for input and output together, so longer text has to be split.
PFN trained the models in two supervised stages on translations it synthesized with Gemma 4, using a tournament that fuses candidate translations, back-translation of naturally written text and iterative self-distillation. It then applied two rounds of preference optimization (DPO), and the models up to 8B were also distilled from the 31B. PFN's own charts compare them with Google's TranslateGemma and Tencent's Hy-MT2 and report that every PLaMo size is ahead on Japanese-English and multilingual tests, that even the 0.8B model beats all of those models in Japanese-English and Japanese-to-other-language translation, including models more than 30 times larger, and that the 2B model is ahead for English-to-other-language translation. The blog calls it the world's best translation-specialized LLM. Those are PFN's benchmarks, including PFTB, an internal Japanese-English set whose references were written with LLMs. On PFTB the 31B model scores 0.6244 and the 0.8B model 0.5862.
Every size comes as BF16 weights, as FP8, W4A16, W4AFP8 and NVFP4 builds for vLLM, as GGUF files for llama.cpp and as 4-bit MLX models for Apple silicon. The 31B model is 63.5 GB in BF16 and 18.7 GB in NVFP4, with a 19.5 GB GGUF at Q4_K_M. The 8B model is 17.1 GB in BF16 and 5.3 GB at Q4_K_M, the 2B model 5.8 GB and 1.8 GB, and the 0.8B model 1.8 GB and 0.57 GB. PFN's tables show most formats within a small margin of BF16, with 2-bit GGUF losing the most. PFN showed the 0.8B model translating on a smartphone in airplane mode in an internal demo and has not released an app. PFN recommends the vLLM Completions API with a hand-built prompt in its PFML3 format for production, and the Transformers chat template, which needs custom code, for trials. The license is the PLaMo Community License, revised on 9 November 2026. It is free for personal and academic use, free for commercial use after registration while revenue or sales stay at or below 1 billion yen, and needs a separate agreement above that. On 11 November 2026 the GGUF repositories were the most downloaded, from about 9,600 downloads in 30 days for the 31B model to about 34,000 for the 0.8B.
Good for
- Self-hosted translation between Japanese, English and the other 51 supported languages
- Translation that has to follow a glossary, extra context or a plain or polite Japanese style
- Offline translation on a laptop with the 0.57 GB 0.8B GGUF or the 1.8 GB 2B GGUF, or on a Mac with MLX
- Translating a long document paragraph by paragraph while passing earlier source and translation pairs as history
Watch out for
- The license is not Apache or MIT, because commercial use needs registration with PFN, is limited to 1 billion yen of revenue including affiliates, and requires the "Built with PLaMo" notice and the PLaMo name in derived models
- The comparisons with TranslateGemma and Hy-MT2 come from PFN's own benchmarks and charts, so test on your own text
- It only translates, and the Japanese plain and polite option works only for Japanese output
- The input and output together are limited to 32,768 tokens, and the Transformers route needs trust_remote_code
- At 2-bit (GGUF Q2_K) accuracy drops, and the 0.8B model loses about 2.5 to 3.3 chrF++ points in the multilingual tests
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the PLaMo 3 Translate 0.8B variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| PLaMo 3 Translate 0.8B BF16 | 890M | 2.05 GB | 1.02 GB | 0.614 GB | Laptop or phone |
| PLaMo 3 Translate 0.8B (GGUF Q4_K_M) INT4 | 890M | 2.05 GB | 1.02 GB | 0.627 GB | Laptop or phone |
| PLaMo 3 Translate 2B BF16 | 2.88B | 6.62 GB | 3.31 GB | 1.99 GB | Laptop or phone |
| PLaMo 3 Translate 2B (GGUF Q4_K_M) INT4 | 2.88B | 6.62 GB | 3.31 GB | 1.99 GB | Laptop or phone |
| PLaMo 3 Translate 8B BF16 | 8.53B | 20 GB | 9.81 GB | 5.89 GB | Laptop or phone |
| PLaMo 3 Translate 8B (GGUF Q4_K_M) INT4 | 8.53B | 20 GB | 9.81 GB | 5.81 GB | Laptop or phone |
| PLaMo 3 Translate 31B BF16 | 31.8B | 73 GB | 37 GB | 22 GB | Gaming GPU |
| PLaMo 3 Translate 31B (NVFP4) NVFP4 | 31.8B | 73 GB | 37 GB | 21 GB | Gaming GPU |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.
The license in plain words
A Japanese community license for the PLaMo models. Personal and academic use is free. Commercial use is free only after you register with Preferred Networks and only while you and your affiliates have at most 1 billion yen of annual revenue or sales; above that you need a separate agreement. If you hand the model or a derived model to others, you must show "Built with PLaMo" and keep "PLaMo" in the names of the models you create.
- Attribution or model name must be shown
- Free commercial use only below a revenue cap
- Acceptable-use policy applies
- Derived models must carry the original name
See also
- PLaMo 3 Translate vs Hy-MT2: which open translation model?
- Best open-source translation models in 2026
- Four open translation models, four different yardsticks
- Free until you grow: where open licenses draw the line
- PLaMo 3 Translate: PFN's open translators for 53 languages
Frequently asked questions
- What is PLaMo 3 Translate?
- PLaMo 3 Translate is a family of AI translation models from the Japanese company Preferred Networks. You pick a size, from a 0.8B model small enough for a laptop CPU to a 31B model for a GPU server. They translate between 53 languages, can follow a glossary and surrounding context, and can write Japanese in a plain or a polite style. The license is free for personal and academic use, but a company has to register first and needs a separate deal above 1 billion yen of revenue.
- Can I use PLaMo 3 Translate commercially?
- Yes, with conditions. PLaMo 3 Translate is released under the PLaMo Community, which allows commercial use but adds extra rules: Attribution or model name must be shown; Free commercial use only below a revenue cap; Acceptable-use policy applies; Derived models must carry the original name. Read the license text before you ship a product.
- Is PLaMo 3 Translate open source?
- PLaMo 3 Translate is open-weight, not open source in the strict sense. You can download the model files, but the PLaMo Community adds conditions that the Open Source Initiative definition does not allow.
- Can I run PLaMo 3 Translate on my own computer?
- Yes. The smallest variant, PLaMo 3 Translate 0.8B, needs about 0.614 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was PLaMo 3 Translate released?
- PLaMo 3 Translate was released by Preferred Networks on November 5, 2026.
- How long a context window does PLaMo 3 Translate have?
- PLaMo 3 Translate supports a context window of up to 32K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- PLaMo 3 Translate 31B model card (Hugging Face) huggingface.co
- PLaMo 3 Translate: World-Class Translation with Open Weights (PFN blog, 5 November 2026) preferred.jp
- PLaMo Community License Agreement (text in the model repository) huggingface.co
- PLaMo 3 Translate 0.8B model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Hy-MT2 Tencent | 2.04B – 30.1B 3B active | 256K | Apache 2.0 (Permissive) | 1.41 GB Laptop or phone | |
| TranslateGemma Google DeepMind | 4.3B – 27.4B | 2K | Gemma Terms (Conditional) | 2.97 GB Laptop or phone | |
| Index-Translate Bilibili | 2.27B – 36B 3B active | 256K | Apache 2.0 (Permissive) | 1.44 GB Laptop or phone | |
| Gemma 4 Google DeepMind | 5.1B – 30.7B | 256K | Apache 2.0 (Permissive) | 3.52 GB Laptop or phone |