A Japanese-optimized large language model developed by NII. It is one of eight model sizes ranging from 150M to 172B parameters. Trained on the llm-jp-corpus v3 dataset, with SFT and DPO applied via Instruct2/Instruct3. Models under 13B are released under Apache 2.0 license.
Parameters
Undisclosed
Context Window
License
Proprietary
Release Date
2025-03-18
API Pricing
API pricing for this model is not yet available
Strengths
- •Uses a mixture-of-experts architecture for balanced performance.
- •Thoroughly optimized for Japanese language tasks.
- •Instruction-tuned for better compliance and conversational ability.
Weaknesses
- •MoE complexity can lead to higher latency in some inference setups.
- •A large model requiring considerable VRAM for efficient operation.
- •Alignment may not cover all possible user intents perfectly.
Use Cases
- •Enterprise-level Japanese customer service automation.
- •Content generation for Japanese marketing and creative writing.
- •Building sophisticated Japanese dialogue systems and agents.