AI News
AI News AgentModel releaseHugging Face3 min read

NVIDIA launches 9B Japanese AI model

NVIDIA has released `Nemotron-Nano-9B-v2-Japanese`, an open model with fewer than 10 billion parameters specialized in Japanese and agent tasks. The company says it leads the under-10B category on the Nejumi Leaderboard 4 and could make private deployments easier for companies.

NVIDIA has released Nemotron-Nano-9B-v2-Japanese, an AI model specialized in Japanese that combines language understanding, reasoning, and tool-use capabilities. It has fewer than 10 billion parameters, a scale that makes it easier to run inside companies and private networks.

The company says the model ranks first among models under 10B on the Nejumi Leaderboard 4, a Japanese evaluation covering nearly 40 tests for language, programming, mathematics, tool use, instruction following, and safety.

A small model for Japanese companies

The goal is to address a specific need: many companies want to use AI with sensitive data without sending it to external services. A model with around 9 billion parameters can be installed more easily on in-house infrastructure than much larger systems.

That does not mean it is a basic model. Nemotron-Nano-9B-v2-Japanese is designed to handle multi-turn conversations, generate code, answer questions, and take actions through tool calls. For example, it can receive a request in Japanese, query an internal API, and return the result in a structured format.

It is based on Nemotron-Nano-9B-v2, a hybrid Transformer and Mamba architecture that NVIDIA had already used to prioritize efficiency. According to company data and research on the base model, this architecture can deliver up to six times more performance than some open-source alternatives under certain conditions. That figure comes from measurements of the Nemotron 2 Nano model, not from a universal promise for every deployment of the Japanese model.

Japanese data and tool-using agents

To adapt it to Japanese, NVIDIA carried out additional training with open corpora such as Wikipedia, FineWeb-2 Japanese, Aozora Bunko, and other text datasets. It also used proprietary data from the Nemotron family, along with tools such as Megatron-LM and NeMo Curator to train and filter the information.

A central component is Nemotron-Personas-Japan, an open dataset with synthetic personas based on demographic, geographic, and personality characteristics from Japan. NVIDIA says the collection was built from six million synthetic people and used as a seed for generating conversation and tool-use examples.

The aim is for the model not only to know how to call a function, but also to understand how to do so in natural conversations and situations common in Japan. This material was used especially to train tasks such as:

  • Querying services and APIs through structured calls.
  • Following instructions in Japanese.
  • Answering questions and solving problems.
  • Maintaining reasoning and programming capabilities.

What changes for developers

The model can be used directly in internal assistants, customer support, process automation, and specialized applications. It can also serve as a starting point for adapting AI to specific sectors such as banking, industry, healthcare, or government, as long as each use case's security and validation requirements are met.

Its size lowers the cost of customization compared with larger models. Instead of training basic Japanese-language and tool-use capabilities from scratch, a team can focus its resources on teaching the model its own terminology, documents, and workflows.

NVIDIA also presents it as an option for building multi-agent systems, meaning applications in which different AI components work together to complete a task. One agent could search for information, another could take an action, and a third could review the result, all within a private infrastructure.

The model is available alongside the Nemotron platform's tools. NVIDIA invites the community to reuse its models, datasets, libraries, and training recipes to create versions adapted to other languages and markets.

The significance of this release is not just that a company has published another 9B model. It is the approach: combining culturally specific data, agent capabilities, and a size compatible with enterprise servers. The next thing to watch is its performance beyond benchmarks: if it maintains that combination in real-world applications, it could become a practical foundation for Japanese-language AI running inside organizations.

NVIDIA launches 9B Japanese AI model | neversleep.ai