Preferred Networks releases open translation AI that runs on phones

Preferred Networks (PFN), the Tokyo AI company, released open weights for PLaMo 3 Translate on 5 November 2026. PLaMo 3 Translate is a family of large language models (LLMs) built only for translation. It comes in four sizes, 0.8B, 2B, 8B and 31B parameters, and is available on Hugging Face. PFN’s main claim is that even the smallest 0.8B model outperforms existing translation-specialised LLMs, including models with more than 30 times as many parameters, in Japanese–English translation and in translation between Japanese and other languages. PFN also says the models are small enough to run fully offline on smartphones, laptops and internal corporate servers.

The quality claims come from PFN’s own testing, which includes proprietary benchmarks. The sources contain no independent evaluation yet.

What PFN released in PLaMo 3 Translate

PFN first developed the models for its own PLaMo Translate service. They are now on the company’s Hugging Face page. All four sizes cover 53 languages, including English and Japanese. The four sizes let developers trade translation quality against latency, processing volume and target hardware.

Each size ships in several formats. PFN provides full-precision BF16 weights, several quantized variants, and files in GGUF and MLX formats. Quantization means storing weights at lower numeric precision to save memory and speed up inference. PFN says it quantized the models with tools such as AutoRound, and that in its benchmarks every quantized format kept nearly the same accuracy as the BF16 version.

Format	Target hardware	Inference engine	PFN’s suggested use
BF16	BF16-compatible GPU	Transformers / vLLM	Unquantized accuracy when memory allows
FP8	NVIDIA Ada Lovelace, Hopper, Blackwell	vLLM	Balance accuracy, memory and speed
W4A16	NVIDIA Ampere or later	vLLM	Less GPU memory with 4-bit weights
W4AFP8	NVIDIA Hopper	vLLM	4-bit weights with FP8 compute for throughput
NVFP4	NVIDIA Blackwell	vLLM	FP4 compute for lower memory and higher speed
GGUF	CPUs / GPUs	llama.cpp	Local runs, optionally splitting across CPU and GPU
MLX 4-bit	Apple silicon	mlx-lm	Running on a Mac with less memory

Source: PFN’s format guide

Beyond plain translation, the models accept surrounding context and glossaries. They also let users choose between plain and polite Japanese. The press release says glossaries can cover proper nouns, technical jargon and product names, and that the tone control for Japanese output covers formal honorific forms and plain forms.

How PFN says it compares with TranslateGemma and Hy-MT2

PFN names its rivals directly. According to its comparative evaluations, the 0.8B version of PLaMo 3 Translate outperforms translation-specific LLMs such as Google’s TranslateGemma and Tencent’s Hy-MT2. The press release includes charts in three categories: Japanese–English, Japanese to multilingual, and English to multilingual. Each chart plots evaluation score against model size on a logarithmic scale, so the best results sit in the top left: small models with high scores. PFN states that these charts are “based on PFN evaluations, including proprietary benchmarks”.

Builders should keep three points in mind:

• The sources give no numeric scores, so this article cannot report margins.
• The strongest wording, beating models with more than 30 times as many parameters, applies to Japanese–English and Japanese-to-other-language pairs. It is not a claim across all 53 languages.
• Some of the benchmarks are proprietary, so outside groups cannot fully reproduce them.

PFN’s technical blog says comparisons against frontier LLM APIs and dedicated translation APIs appear in an earlier, Japanese-language release post. The sources here do not include that post, and the sources do not show how PLaMo 3 Translate compares with earlier PLaMo translation models.

How PFN built the models

PLaMo 3 Translate starts from PLaMo 3 NICT. That base model was pretrained through joint research between PFN and Japan’s National Institute of Information and Communications Technology (NICT). PFN describes the starting point as a third-generation PLaMo base model that had finished pretraining but had not been fine-tuned.

Training then runs in three stages:

1. Phase 1 supervised fine-tuning (SFT). The model trains on a large volume of (source text, translation) pairs to learn the basic translation task.
2. Phase 2 SFT. Training uses the same kind of pairs, but with more varied source texts that better reflect real-world use. The synthetic translations are refined more thoroughly. The goal is to move from large-scale data to a smaller, higher-quality set.
3. Direct preference optimization (DPO). For each source text, PFN picks a preferred (“chosen”) translation and a weaker (“rejected”) one from several candidates. It then trains the model to give the chosen version higher probability.

For the 8B and smaller models, PFN adds a distillation step between SFT and DPO. Distillation means training a smaller model to imitate a larger one, here the 31B model. PFN credits the results to careful data refinement, using techniques such as tournament-based data synthesis, back-translation and iterative self-distillation, combined with distillation into the smaller models. Back-translation means translating text back into its original language to create or check training pairs.

Licence terms and who pays

The weights are released under the PLaMo Community License, PFN’s own licence, not a standard open-source one. According to PFN’s technical blog:

• Private use, such as personal study and hobbies, is free.
• Noncommercial use, including academic use, has no income or revenue limit and needs no registration with PFN.
• Commercial use is free if the user or its affiliates had income or revenue of JPY 1 billion or less in the most recent fiscal year, provided the user registers in advance through a dedicated form.
• Commercial use above that threshold requires a separate licence agreement with PFN.

The press release states the same threshold, giving it as approximately $6.5 million, and says commercial use includes internal corporate operations and integration into products or services. The wording is not fully consistent. The release body says “1 billion Japanese yen or less” in one place and “less than 1 billion Japanese yen” in another. A correction note added at 11:45 on 5 November also rewrites the threshold sentence. Companies near the line should read the licence text on the Hugging Face model pages before relying on the free tier.

What it means for builders

The release mainly matters for teams that need translation without sending text to a cloud service. PFN argues that because the model runs on user-managed devices or servers, neither the source text nor the translation goes to external cloud servers. It names contracts, R&D documents and customer support logs as typical sensitive material.

PFN demonstrates the 0.8B model translating on a smartphone in airplane mode. It describes the demo as running at normal speed in a demo environment. The technical blog adds a caveat: this is an internal demo, and PFN has not released an app like it. In practice, developers get model files and must build the mobile or desktop integration themselves. The GGUF and MLX builds point to llama.cpp and mlx-lm as the likely routes for laptops and Macs. PFN expects uses in tourism, enterprise business and disaster response. It also lists wearable devices among the targets for the smaller models.

The release also changes how teams can set up translation. Some existing tools that call Hugging Face for translation expect one model per language pair. Pimcore’s documentation for its Copilot translation action, for example, tells users to pick a Helsinki-NLP opus-mt model named for the input and output language, such as an English-to-French model, and to match the configured endpoint to that pair. PLaMo 3 Translate is one model covering 53 languages, which could reduce the number of models a multilingual product has to manage. The sources do not say whether PLaMo 3 Translate works with Pimcore or similar tools.

For the broader trend of small models built to run offline, see Unicorn API’s earlier coverage of Soket Labs and ZenteiQ’s offline Indian AI models.

Limits and open questions

• No independent benchmarks yet. Every quality comparison in the sources is PFN’s own, partly on proprietary tests.
• Uneven language evidence. The headline claims focus on Japanese-related pairs and English to multilingual. The sources give no breakdown for each of the 53 languages.
• Quantization claims are self-reported. PFN says quantized versions keep nearly BF16 accuracy and refers readers to the model cards for details. Teams should check results on their own phones and laptops, since quantized models can behave differently across devices.
• Licence ambiguity. The corrected and inconsistent threshold wording in the press release, and the registration step described in the blog, add work for commercial users.
• Long inputs. The sources do not state a context window or any limit on input length.

What to watch next

The first test is independent evaluation. Translation benchmarks and community testers can now check PFN’s claims against TranslateGemma and Hy-MT2 using public data. A second signal is whether PFN, or developers building on the open weights, ship an actual offline mobile translation app, since PFN’s phone demo is internal for now. PFN says the release is meant to raise awareness of its PLaMo foundation models and to encourage open innovation around LLMs on edge devices. Uptake of the four sizes, and whether larger companies sign commercial licences, will show how far that goes.