PLaMo 3 Translate: PFN's open-weight translation LLM runs offline on a phone in 53 languages

Preferred Networks announced the release on November 5, 2026: the weights for PLaMo 3 Translate, the translation-specialized large language model behind its PLaMo Translate service, are now public on Hugging Face. The suite comes in four sizes — 0.8B, 2B, 8B and 31B parameters — and is built to run natively on the user’s own hardware, from a smartphone to an on-premises server, with no internet connection at all.

The pitch is the opposite of the cloud-translation default. Instead of sending text to an API, you download the weights and translate locally. PFN demonstrated the smallest 0.8B model translating in real time on a smartphone in airplane mode, and says the tiered sizes let developers trade translation quality against latency, volume and hardware — a phone gets the 0.8B model, an enterprise server the 31B one.

Four sizes, from phone to server

All four models cover the same 53 languages, including English and Japanese, and all four ship in multiple formats: BF16 weights plus FP8, W4A16, W4AFP8 and NVFP4 quantized variants, and GGUF and MLX builds for CPU, Apple silicon and local inference engines. PFN quantized the models with tools including AutoRound and says every quantized format retains nearly the same accuracy as the unquantized BF16 version.

The models are built on PLaMo 3 NICT, a pre-trained model developed through joint research between PFN and Japan’s National Institute of Information and Communications Technology (NICT). Training ran through two stages of supervised fine-tuning followed by direct preference optimization, with the smaller models distilled from the larger ones. PFN’s tech blog puts the training data at about 37.3 billion tokens, refined through tournament-based data synthesis, back-translation and iterative self-distillation.

The small model is the story

PFN’s headline claim is about the smallest model, not the largest. According to the company’s comparative evaluations, the 0.8B version outperforms comparable translation-specific LLMs such as Google’s TranslateGemma and Tencent’s Hy-MT2 — and, per PFN’s tech blog, beats existing translation-specialized models with more than 30 times as many parameters in Japanese–English and Japanese-to-other-language translation. The company calls PLaMo 3 Translate “the world’s best translation-specialized LLM” on the strength of those results, which come from its own benchmarks, including a proprietary Japanese–English benchmark.

The release lands in a week crowded with open-weight drops. Baidu put out ERNIE 5.5 Flash Base, a 280-billion-parameter model for fine-tuning, and Tencent released Hy4, a 770-billion-parameter model with vision and computer use. PLaMo 3 Translate is the outlier of the two: a specialist, not a generalist, and the only one designed to run on a phone.

Privacy, glossaries and tone control

Running locally is also a privacy argument. Because neither the source text nor the translation ever leaves the user’s device or server, PFN positions the models for contracts, R&D documentation and customer-support logs — material that companies are reluctant to send to a cloud translation API. The company points to tourism, enterprise business and disaster response as the use cases it expects the release to unlock.

The models also take surrounding context and custom glossaries, so proper nouns, technical jargon and product names translate consistently, and when translating into Japanese users can specify formal honorific forms or plain language. That control is aimed squarely at professional translation workflows, where a mistranslated product name or the wrong level of politeness is a defect, not a quirk.

The license: free up to ¥1 billion in revenue

Under the PLaMo Community License, individuals and enterprises with annual revenue of 1 billion Japanese yen or less in their most recent fiscal year — about $6.5 million — can download and use the models free of charge, for commercial and non-commercial applications alike. Organizations above that threshold need a commercial license agreement with PFN. The weights are available now on PFN’s Hugging Face page, in all four sizes.

For a Japanese lab whose foundation model, PLaMo, is far less known outside Japan than its rivals, the release is also a visibility play: PFN says it wants to expand recognition of PLaMo while accelerating open innovation in edge-device language models. A translation model that fits in a pocket is a good way to be noticed.