Verdict: for offline translation involving Japanese, or on a phone or small laptop, PLaMo 3 Translate is the one to try first. Preferred Networks (PFN) released the PLaMo 3 Translate open weights on 5 November 2026 in four sizes (0.8B, 2B, 8B and 31B parameters) covering 53 languages, and its own benchmarks place it ahead of Google DeepMind's TranslateGemma and Tencent's Hy-MT2. Those scores are vendor-run and no independent evaluation exists yet, so for language pairs that do not involve Japanese, treat the lead as a reason to test rather than a settled result.
TL;DR
- PLaMo 3 Translate is a translation-only model with open weights on Hugging Face, released 5 November 2026 in 0.8B, 2B, 8B and 31B sizes.
- It supports 53 languages and runs fully offline; PFN demonstrated the 0.8B model on a smartphone in airplane mode.
- PFN's benchmarks say the 0.8B model beats translation models more than 30 times its size on Japanese pairs. These are vendor figures, not independent tests.
- Commercial use is free only after registration, and only for organisations at or under JPY 1 billion in annual revenue.
- Official GGUF and MLX files exist, so no third-party conversion is needed.
- Last verified: 8 November 2026.
What do PLaMo 3 Translate open weights include?
PFN publishes each of the four sizes in several formats: BF16, FP8, W4A16, W4AFP8, NVFP4, GGUF and MLX-4bit. GPU formats run through Transformers or vLLM, GGUF runs in llama.cpp on CPU or GPU, and MLX runs on Apple silicon through mlx-lm. According to PFN's release notes, its benchmarks show the quantised formats keeping nearly the same accuracy as BF16.
The model accepts surrounding context and glossaries, and for Japanese output you can choose plain (常体) or polite (敬体) style. It is not a chat model, and the maximum training length is 32,768 tokens for input and output combined, so long documents need splitting.
How big are the files, and what hardware do you need?
Size is the practical constraint for offline use. Measured Q4_K_M GGUF sizes are 0.53 GiB for 0.8B, 1.68 GiB for 2B, 4.92 GiB for 8B and 18.19 GiB for 31B.
| Size | Q4_K_M GGUF file | Runs on | Serve with |
|---|---|---|---|
| 0.8B | 0.53 GiB | Phones, older laptops, CPU-only machines | llama.cpp |
| 2B | 1.68 GiB | Light laptops, CPU-only machines | llama.cpp |
| 8B | 4.92 GiB | One consumer GPU, or a Mac with MLX-4bit | vLLM or llama.cpp |
| 31B | 18.19 GiB | An on-premises server | vLLM FP8 |
If you already run local models for other work, such as the coding stacks compared in our local LLM coding comparison, the same machine will handle the 8B build comfortably. For a sense of how local inference speed varies by runtime, see our Gemma 4 speed comparison.
Does PLaMo 3 Translate beat TranslateGemma and Hy-MT2?
On PFN's numbers, yes. PFN reports that all four models outperform existing translation-specialised LLMs on Japanese-English and multilingual tests, that the 0.8B model surpasses models more than 30 times its size on Japanese-English and Japanese-to-other-language pairs, and that the 2B model leads all compared models on English-to-other-language pairs. The named comparison models are TranslateGemma (arXiv 2601.09012) and Hy-MT2 (arXiv 2605.22064). The 0.8B model also beats PFN's own 8B PLaMo 2 Translate.
The caveats matter. The tests include PFN's in-house PFTB and SaaS benchmarks alongside public sets (WMT24++ scored with COMET-22, and a 15-language subset of FLORES+ scored with chrF++). Proprietary benchmarks cannot be reproduced, and independent coverage notes that no third-party evaluation has been published. Run a few hundred sentences from your own domain before committing.
How was PLaMo 3 Translate trained?
PFN's technical write-up describes a base model, PLaMo 3 NICT, built jointly with Japan's NICT. Supervised fine-tuning ran in two stages: a larger mined and synthesised set, then a smaller curated set, for a total of about 37.3B tokens. Preference tuning with DPO followed in two rounds, and the 8B and smaller models were distilled from the 31B teacher.
Two details stand out. Data synthesis used Google's open Gemma 4 31B IT, with back-translation and tournament-style selection. And PFN tried GRPO reinforcement learning and rejected it, because human evaluators found fluency declined. In the 31B DPO tests, judged by Gemini 3.1 Pro, the first DPO round beat the SFT model 57% of the time for English to Japanese and 59% for Japanese to English, according to PFN.
How do you run it offline?
- Choose a size from the hardware list above.
- On a Mac, use mlx-lm with the
pfnet/plamo-3-translate-8b-MLX-4bitrepository. On Windows or Linux CPU, use llama.cpp with PFN's official GGUF repositories. On a GPU server, serve an FP8 or W4A16 build with vLLM. - Pass the target language and style through the chat template, for example
{"target_lang": "ja", "form": "plain"}, and set temperature to 0. - Put context and glossary terms in the prompt, and keep input plus output under 32,768 tokens.
# Illustrative: check exact repository names on the pfnet Hugging Face organisation
vllm serve pfnet/plamo-3-translate-8b-FP8
If you are upgrading from PLaMo 2 Translate, the new 0.8B model beats the old 8B on PFN's figures, so you can often drop a size class and keep quality.
What does the PLaMo Community License allow?
Personal and academic use is free, with no revenue limit and no registration. Commercial use is free if you and your affiliates had annual revenue of JPY 1 billion or less in the latest fiscal year, and you register in advance through PFN's form. Above that, you need a separate commercial licence from PFN. PFN's press release puts the threshold at about $6.5 million, or roughly ₹56 crore at current rates; the binding figure is the yen amount in the PLaMo Community License.
PFN corrected the wording on launch day from "less than" to "1 billion yen or less". Companies close to the line should read the licence text rather than a summary.
Which should you choose?
Choose PLaMo 3 Translate if Japanese is one of your languages, if you need a model small enough for a phone, or if confidential contracts and medical documents must stay on your own hardware. Consider TranslateGemma or Hy-MT2 if your pairs exclude Japanese and you want a model already tested by people outside its maker, or if the registration and revenue terms do not suit you. Other recent open-weight releases, such as Kimi K3.1 Pro and Qwen 3.8 Max, are general-purpose models and a different kind of choice. Why we covered this one first: a scan of our live article corpus (n=435 slugs via the public article API, run 2026-11-08) found no prior coverage of any translation-specialized open-weight model, making this the first on the site.
FAQ
Q: Are PLaMo 3 Translate open weights free for commercial use?
A: Yes, if your organisation and its affiliates had annual revenue of JPY 1 billion or less in the most recent fiscal year and you register in advance with PFN. Larger organisations need a separate commercial licence from PFN.
Q: Which languages does PLaMo 3 Translate support?
A: It supports 53 languages, including English and Japanese. PFN's public FLORES+ comparison covers a 15-language subset that includes Arabic, Korean, Thai, Vietnamese, German and French.
Q: Can PLaMo 3 Translate run on a phone or laptop?
A: Yes. The 0.8B Q4_K_M GGUF file is 0.53 GiB and the 2B file is 1.68 GiB, and PFN demonstrated the 0.8B model translating on a smartphone in airplane mode.
Q: Is PLaMo 3 Translate a general chatbot?
A: No. It is a translation-only model, so it is not suited to open-ended chat, and combined input and output is limited to 32,768 tokens.
Q: Are the benchmark results independent?
A: No. The results come from PFN's own tests, including proprietary benchmarks, and no independent evaluation has been published yet, so test it on your own text before relying on it.
Sources: PFN press release, PFN technical blog, PLaMo Community License, toolhalla.ai local AI roundup. Last verified 8 November 2026.
Top comments (0)