Skip to main content
Rase AI Intelligence Newsletter

Models & assistants Model or AI service

LLM-jp-3 offers downloadable Japanese language models for research

Japan’s National Institute of Informatics released LLM-jp-3 models built on a Japanese-oriented training corpus. The initial public group included 1.8B, 3.7B and 13B models. [2][3]

Model access
Downloadable base and instruction checkpoints within the LLM-jp-3 series. [1][3]
License
Apache License 2.0 for the cited 13B base and the initial 1.8B, 3.7B and 13B releases. [2][3]

What the model does

The center announced those first sizes in September 2024. It later added more sizes and published new instruction-tuned checkpoints using the same series. [1][2][3]

Why it matters

Making checkpoints and development information public gives researchers more material for reproducing Japanese-language experiments. Several model sizes also allow comparisons under different hardware limits. [2][3]

Who can use it

Researchers working on Japanese text can use the 13B base as a starting point for studies or further training. Developers can evaluate a separate Instruct version for conversational tasks. [1][3]

Key dates

  1. LLM-jp-3 1.8B, 3.7B and 13B released [2]
  2. Additional Instruct2 and Instruct3 checkpoints announced [1]

Ways to use it

Compare a checkpoint on a representative Japanese text task with one or more sizes of the same series. Record whether it is a base model or an instruction-tuned version. [1][3]

Model limitations

This release spans distinct model sizes and versions with different terms. The 13B weights cited here are under Apache 2.0, and the later 172B variant has separate licensing. [1][3]

Sources and verification

Catalog evidence checked How we verify records

  1. Primary source NII’s LLM-jp-3 Instruct2 and Instruct3 release and variant license details
  2. Primary source NII announcement of LLM-jp-3 1.8B, 3.7B and 13B
  3. Primary source llm-jp/llm-jp-3-13b
  4. Primary source Hugging Face registry metadata for llm-jp/llm-jp-3-13b

Related profiles

Models & assistants

GPT-6.2 Astra and Sol

OpenAI

Astra targets demanding projects. Sol targets everyday tasks. Both support text and image input in the API.

High confidenceEvidence checked

Models & assistants

MAI Code 2 and Code 2 Flash

Microsoft AI

Code 2 targets complex software work. Flash targets frequent edits, explanations and routine fixes.

High confidenceEvidence checked

Models & assistants

MAI Realtime 1

Microsoft AI

Users can try a MAI Labs demo. Developers can register for a limited Foundry preview.

High confidenceEvidence checked