llm-jp/awesome-japanese-llm

What it solves

Information about Japanese-focused large language models (LLMs) is scattered across company announcements, research papers, and model hubs, making it hard to compare options. This project collects that information into one community-maintained summary of publicly available Japanese LLMs (models trained mainly on Japanese) and Japanese LLM evaluation benchmarks, so readers can see key facts side by side instead of digging through many sources.

How it works

Volunteers gather information — some of it cited from papers and public resources — and organize it into structured tables on GitHub. For each model, the tables list the release year, architecture (such as Llama or Mixture-of-Experts), how many tokens the model can handle in input and output, the training text used, the developer, and the license or terms of use. Models are grouped by purpose: the main section covers text generation models, with a dedicated subsection for general-purpose models trained from scratch, and a separate section is referenced for models that also handle images. The project also flags an important caveat: while many models use open-source licenses like MIT or Apache-2.0, some use non-commercial or developer-specific licenses (for example CC BY-NC-SA 4.0) and may not truly be open source. A more readable web version is available, and corrections or suggestions for new models can be submitted through GitHub Issues.

Who it's for

Researchers, developers, and companies that need to discover, compare, or select a Japanese-language LLM — especially anyone who wants a quick structured overview of model capabilities and licensing before adopting or building on a model.

Highlights

• Covers both Japanese LLMs and Japanese LLM evaluation benchmarks in one place
• Lists models from many developers, including SB Intuitions, NII / the large language model research center, Stockmark, Preferred Networks, CyberAgent, and ELYZA
• Each entry records release year, architecture, context token length, training data, developer, and license
• Spans a wide range of sizes from roughly 150M up to 172B parameters, including Mixture-of-Experts models
• Explicitly warns which publicly released models are not genuinely open source due to non-commercial or developer-specific licenses
• Provides newer models with long context windows (up to 65,536 tokens)
• Maintained by volunteers on GitHub, with community contributions handled through GitHub Issues
• Offers a more readable web version at llm-jp.github.io/awesome-japanese-llm