Japanese AI Models: A Practical Guide

A overview of notable large language models developed in Japan

Japan’s AI ecosystem has grown quickly, driven by a simple problem: global models handle Japanese well enough, but often miss its nuance, honorifics, mixed scripts and cultural context. Universities, tech giants, telecoms and startups have responded with models tuned for Japanese language, enterprise privacy and domestic compute. Some are open for anyone to download; others are sold to businesses. This guide lists the most notable ones and breaks down what each does best. Model names and versions change fast, so confirm current details with each developer before building on them.

Overview Table

Model	Developer	Access	Main Focus
Fugaku-LLM	Fujitsu, RIKEN, Tokyo Tech and partners	Open	Japanese LLM trained on a supercomputer
LLM-jp	National Institute of Informatics (NII)	Open	Transparent, research-grade Japanese LLMs
Swallow	Institute of Science Tokyo and AIST	Open	Japanese-boosted versions of open models
PLaMo	Preferred Networks	Open and commercial	Efficient, domestically built foundation models
Sarashina	SB Intuitions (SoftBank)	Open and commercial	Large-scale Japanese models
tsuzumi	NTT	Commercial	Lightweight enterprise LLM
cotomi	NEC	Commercial	Fast, secure business LLM
Takane	Fujitsu (with Cohere)	Commercial	Enterprise LLM for regulated sectors
Rakuten AI	Rakuten	Open and commercial	Japanese models for e-commerce and services
ELYZA models	ELYZA (KDDI group)	Open and commercial	Japanese-tuned assistants
Sakana AI research	Sakana AI	Mostly research	Nature-inspired, efficient AI methods

Model Breakdown

Fugaku-LLM

Trained on the Fugaku supercomputer, this model showed that Japan could build a large language model on domestic CPU-based infrastructure rather than relying on GPU clusters abroad.

Features

• Trained from scratch with a heavy share of Japanese data
• Released openly for research and commercial use
• Built through a multi-institution collaboration

Strengths

• Strong Japanese language handling for its size
• Proof of concept for sovereign compute
• Useful baseline for researchers

LLM-jp

Led by the National Institute of Informatics, LLM-jp is an open project that publishes models along with their data and training details.

Features

• Multiple model sizes released over time
• Open corpora, tooling and evaluation resources
• Community of academic and industry contributors

Strengths

• Transparency that few commercial models offer
• Ideal for studying how Japanese LLMs are built
• Free to adapt and fine-tune

Swallow

Swallow takes strong open models, such as Llama-family models, and continues their training on large amounts of Japanese text.

Features

• Continual pretraining on Japanese corpora
• Instruction-tuned variants available
• Regular updates tracking new base models

Strengths

• Excellent Japanese performance without training from scratch
• Practical for developers who want self-hosted models
• Keeps English and reasoning ability largely intact

PLaMo

From Preferred Networks, PLaMo is a family of foundation models built in-house, with attention to efficiency and architecture.

Features

• Models for both Japanese and English
• Compact variants suited to edge and cost-sensitive use
• Both open releases and commercial offerings

Strengths

• Strong engineering heritage in deep learning
• Good performance per unit of compute
• Suited to industrial and enterprise applications

Sarashina

Sarashina comes from SB Intuitions, the SoftBank subsidiary focused on Japanese AI infrastructure and services.

Features

• Several parameter scales, including very large models
• Trained on extensive Japanese web and curated data
• Backed by substantial domestic compute investment

Strengths

• Natural, fluent Japanese output
• Corporate backing and integration with SoftBank services
• Scale ambitions that few Japanese players match

tsuzumi

NTT’s tsuzumi is designed to be small and efficient, so organizations can run it on modest hardware.

Features

• Lightweight architecture with low running cost
• Strong Japanese and English capability
• Options for tuning to company-specific data

Strengths

• Lower infrastructure requirements
• Appeals to privacy-conscious enterprises
• Backed by NTT’s research depth and customer network

cotomi

NEC’s cotomi targets business use, emphasizing speed and secure deployment.

Features

• Fast inference suitable for real-time work
• Deployment on private or on-premises environments
• Domain customization for corporate documents

Strengths

• Fits strict data governance needs
• Good for internal knowledge search and document tasks
• Strong enterprise sales and support channels

Takane

Fujitsu’s Takane is an enterprise model built with Cohere, aimed at sectors such as finance, government and healthcare.

Features

• Strong Japanese language and document comprehension
• Private deployment options
• Tools for tailoring to specialist vocabularies

Strengths

• Designed for regulated industries
• Reliable handling of formal and technical Japanese
• Supported by Fujitsu’s integration services

Rakuten AI

Rakuten has released Japanese-focused models, some open, and uses them across its e-commerce, finance and telecom services.

Features

• Japanese-optimized models of various sizes
• Built on open foundations and extended in-house
• Integrated into Rakuten’s own products

Strengths

• Real-world tuning from huge volumes of consumer interactions
• Strong fit for retail and customer service tasks
• Open options for developers

ELYZA

A startup spun out of the University of Tokyo’s Matsuo Lab, now part of the KDDI group, ELYZA builds Japanese-tuned assistants.

Features

• Japanese instruction-tuned models, including open releases
• Focus on summarization, writing and business tasks
• Enterprise tooling around the models

Strengths

• Polished Japanese writing quality
• Practical, task-oriented tuning
• Telecom backing for distribution

Sakana AI

Tokyo-based Sakana AI is known less for one flagship model than for research ideas, such as evolutionary model merging and automated scientific discovery.

Features

• Methods for combining existing models into new ones
• Experiments in AI-driven research workflows
• Small, efficient model research

Strengths

• Creative alternatives to brute-force scaling
• Influential in the global research conversation
• Interesting for teams exploring efficiency

Choosing a Model

The right choice depends on your goal. For research and learning, LLM-jp and Fugaku-LLM offer the most transparency. For self-hosting with strong Japanese output, Swallow and open PLaMo or Sarashina releases are practical. For enterprise deployment with support and security, tsuzumi, cotomi and Takane are built for that world. For consumer and retail applications, Rakuten AI and ELYZA are worth testing.

Whatever you pick, test on your own Japanese content, including keigo, industry jargon and mixed kanji and kana text. Benchmarks help, but your own documents are the better judge.