Large Language Models (LLMs) and ChatGPT-related
10.1 LLM Development Projects
In recent years, the development of large language models supporting Japanese has been increasingly active, with multiple large-scale projects underway.
LLM-jp project: Outcomes of the LLM study group centered on the National Institute of Informatics (NII).
- llm-jp-eval (Stars: 175): LLM automated evaluation tool
- llm-jp-sft (Stars: 66): Supervised fine-tuning
- llm-jp-tokenizer (Stars: 54): LLM tokenizer
- llm-jp-corpus (Stars: 49): Training corpus
- llm-jp-eval-mm (Stars: 44): Multimodal LLM evaluation
- llm-jp-judge (Stars: 64): Automated generation evaluation
- llm-jp-vila (Stars: 10): VILA-based VLM training
Swallow Project: A Japanese LLM development project centered on Tokyo Institute of Technology.
- swallow-evaluation (Stars: 26): Evaluation scripts
- swallow-evaluation-instruct (Stars: 34): Post-training model evaluation
- swallow-corpus (Stars not listed): Large-scale Japanese web corpus
shisa-v2 (shisa-ai/shisa-v2): A Japanese/English bilingual LLM. Stars: 35.
OneCompression (FujitsuResearch/OneCompression, Stars: 445): A post-training quantization (PTQ) pipeline for LLMs developed by Fujitsu Laboratories. It integrates QEP (NeurIPS 2025), ILP mixed precision, rotation preprocessing, and a vLLM plugin — a comprehensive framework for efficient local LLM inference.
Japanese LLMs on Hugging Face
In the text-generation category on Hugging Face, numerous Japanese LLMs have been published. The models with the highest download counts are shown below.
| Model | HF DL | Likes | Developer | Description |
|---|---|---|---|---|
| llm-jp-4.1-8b-thinking | 80k | 18 | NII/LLM-jp | 8B reasoning model from the llm-jp-4.1 series |
| llm-jp-4-8b-instruct | 49k | 16 | NII/LLM-jp | Instruction-tuned 8B model from llm-jp-4 series |
| sarashina2.2-3b-instruct-v0.1 | 16k | 39 | SB Intuitions | Autoregressive Japanese model |
| Llama-3-Swallow-8B-Instruct-v0.1 | 9k | 21 | Tokyo Institute of Technology | Meta Llama 3 Japanese extension |
| Llama-3.1-Swallow-8B-Instruct-v0.5 | 7k | 20 | Tokyo Institute of Technology | Trained on synthetic Japanese data |
| Mistral-Nemo-Japanese-Instruct-2408 | 6k | 49 | CyberAgent | Japanese version of Mistral-Nemo |
| LFM2.5-1.2B-JP-202606 | 4k | 82 | Liquid AI | Hybrid architecture |
| Llama-3-ELYZA-JP-8B | 4k | 150 | ELYZA | High popularity with 150 Likes |
| llm-jp-4-8b-thinking | 3k | 52 | NII/LLM-jp | 8B model with reasoning trace generation |
| NVIDIA-Nemotron-Nano-9B-v2-Japanese | 3k | 146 | NVIDIA | Mamba-2/MLP/Attention mixture |
| llm-jp-4-32b-a3b-thinking | 2k | 40 | NII/LLM-jp | 32B MoE reasoning model |
| TinySwallow-1.5B-Instruct | 2k | 61 | Sakana AI | Lightweight model via TAID distillation |
The shift toward reasoning (thinking) models continues. The newly released llm-jp-4.1-8b-thinking (80k/week) now tops this table, and the GGUF build llm-jp-4-33b-thinking-gguf (271k/week, #10 across all Japanese NLP models on Hugging Face) has also grown sharply. Meanwhile, the previous leaders llm-jp-4-8b-thinking (21k→3k/week) and llm-jp-4-32b-a3b-thinking (16k→2k/week) declined as the 4.1 series arrived, while the instruction-tuned llm-jp-4-8b-instruct (49k/week) holds steady. NVIDIA-Nemotron-Nano-9B-v2-Japanese stays at 3k/week, though its 146 Likes remain high.
The Swallow project is also extremely active on Hugging Face, extending diverse base models such as Llama 3/3.1, Gemma 2, GPT-OSS, and Qwen3 to Japanese. In particular, GPT-OSS-Swallow provides large-scale models of 20B and 120B parameters, reinforcing Japanese capabilities while maintaining math and coding performance.
The ELYZA series receives strong community support with high Like counts (Llama-3-ELYZA-JP-8B: 150 Likes, ELYZA-japanese-Llama-2-7b-instruct: 75 Likes).
TinySwallow-1.5B (Sakana AI) is a model distilled from Qwen2.5-32B-Instruct using TAID distillation technology, demonstrating the potential of compact yet high-performance Japanese models.
LFM2.5-1.2B-JP (Liquid AI): Adopts a hybrid architecture combining Mamba-2 and MLP, achieving JMMLU 50.7, M-IFEval 58.1, and GSM8K 56.0. The June 2026 release (LFM2.5-1.2B-JP-202606) reaches 4k/week (82 Likes) and is also distributed in ONNX and GGUF formats; the GGUF build sees 8k/week, more than the original.
10.2 ChatGPT/LLM Applications
Japanese applications of ChatGPT and LLMs span a wide range.
Dialogue / Chatbots:
- ChatdollKit (Stars: 1,300): A Unity toolkit for turning 3D models into chatbots
- aiavatarkit (Stars: 704): Building AI-based conversational avatars
- chatvrm (Stars: 848): A VRM character conversation demo by pixiv
VTubers / Virtual Characters:
- gptuber-by-langchain (Stars: 63): A system where GPT acts as a YouTuber
- VRChatGPT (Stars: 15): Conversation on VRChat
- AITuberDegikkoMirii (Stars: 5): AITuber platform
RAG (Retrieval Augmented Generation):
- RAG-Japanese / rag-japanese (Stars: 10): RAG for Japanese LLMs
- azure-search-openai-demo (Stars: 46): Japanese implementation of the RAG pattern
- wikipedia-japanese-open-rag: RAG with Japanese Wikipedia
Prompt Engineering:
- chatgpt-prompt-sample-japanese (Stars: 431): Japanese prompt samples
10.3 Evaluation Benchmarks
- IgakuQA (Stars: 52): GPT-4/ChatGPT evaluation on the Japanese medical licensing exam
- Japanese-RP-Bench: Japanese role-playing ability benchmark
- japanese-llm-roleplay-benchmark (Stars: 42): LLM character role-play evaluation
- pfmt-bench-fin-ja (Stars: 9): Japanese financial benchmark
- igakuqa119 (Stars: 10): LLM evaluation on the 119th Japanese medical licensing exam
- jmle2026-bench (Stars: 11): LLM benchmark on the 120th Japanese Medical Licensing Examination (Feb 2026)
- japanese-hr-niah (kufu): Long-context LLM evaluation in the Japanese HR/labor domain using Needle-In-A-Haystack format
- Doppelganger-JC: Benchmark for cross-lingual homograph misuse between Chinese and Japanese in LLMs
- JMultiWOZ-TC (llm-jp): Agent Function Calling evaluation in multi-turn dialogue scenarios