← Back
NVIDIA releases Nemotron-Nano-9B-v2-Japanese, tops sub-10B category on Nejumi Leaderboard
· releasemodelfeature · huggingface.co ↗

Overview

NVIDIA has released Nemotron-Nano-9B-v2-Japanese, a Japanese-specialized variant of its efficient small language model. The model achieves top performance on the Nejumi Leaderboard 4 in the sub-10B parameter category, addressing a critical gap in the Japanese enterprise AI market where few models combine both advanced Japanese proficiency and agentic task execution capabilities.

Key Features & Capabilities

The model maintains the proven architecture of the English-optimized Nemotron-Nano-9B-v2 while adding specialized Japanese language capabilities. Notable capabilities include:

  • Advanced Japanese language understanding with strong reasoning and knowledge retention
  • Robust tool-calling and agent abilities for function invocation and multi-step workflows
  • Instruction-following and alignment across multiple dimensions
  • Efficient parameter count (9B) enabling on-premises deployment without significant infrastructure investment

Training Approach

The development leverages two foundational pillars:

Nemotron-Nano-9B-v2 Architecture: A proven baseline with excellent size-to-performance ratio, adapted from the English-trained model.

Nemotron-Personas-Japan Dataset: An open-source (CC BY 4.0) dataset of 6M demographically and geographically representative Japanese personas used as seed data for synthetic data generation. This culturally-grounded approach ensures training data reflects real-world Japanese diversity and use cases, particularly for tool-calling scenarios.

The training pipeline combines:

  • Continued pretraining on Japanese OSS corpora (Wikipedia, fineweb-2-ja, aozorabunko, sip3-ja-general-web-corpus)
  • Synthetic data generation (SDG) using Nemotron-Personas-Japan as seed set
  • Supervised fine-tuning (SFT) with tool-calling and instruction-following data
  • Training uses Megatron-LM and NeMo Curator tools

Enterprise Relevance

The model addresses specific Japanese enterprise needs:

  • On-premises deployment: Sub-10B size enables private network deployment for organizations handling sensitive data
  • Customization efficiency: Strong base capabilities reduce fine-tuning cycles compared to foundation models
  • Agent development: Proven agent architecture supports rapid prototyping of multi-agent systems without large-model overhead

Availability

The model is available on Hugging Face at nvidia/NVIDIA-Nemotron-Nano-9B-v2-Japanese. Related datasets including Nemotron-Personas for other regions (US, India, Singapore, Brazil) are also available, enabling similar customization approaches across markets.