NTT Boosts Japanese AI Performance with Small LLM tsuzumi 2

Summary

NTT's tsuzumi 2 model is a 28-billion-parameter dense Large Language Model (LLM) developed from scratch. The model is optimized for Japanese business processes and inference on a single GPU. It is positioned as an alternative that ensures data sovereignty against larger foreign models. tsuzumi 2 features a custom morphological tokenizer, pre-training on over 10 trillion tokens, and strong Japanese instruction-following scores on M-ifeval.

Why it matters

For Japanese companies, the ability to run a capable LLM on-premises without exposing internal data to foreign cloud providers is a practical and legal necessity. Larger US- and EU-based models do not fully meet this need. The 28B dense architecture, which can be deployed on a single GPU, makes the model genuinely usable in mid-sized enterprise environments without requiring hyperscale infrastructure. NTT's emphasis on token-efficient Japanese processing directly reduces inference costs in high-volume workflows such as document summarization and meeting note generation.

Highlights

• 28B dense (non-MoE) model designed to fit on a single GPU for on-premises inference
• Custom Japanese tokenizer based on decades of morphological analysis research to reduce token count and cost
• Trained on over 10 trillion tokens; outperformed comparison models in Japanese instruction following (M-ifeval) and format adherence tasks
• Training infrastructure: AWS H100 clusters + SageMaker HyperPod for stable and fault-tolerant fine-tuning cycles
• Related research: VDocRAG (visual document RAG), ToMATO sentiment understanding benchmark, Portable Reward Tuning, Wavelet Positional Encoding

Critical take

This text reads like a restructured transcript of an AWS Summit keynote. Consequently, it comes across as vendor-centric marketing content rather than an independent analysis. The text lacks third-party benchmarks, publicly available model weights, or API access. No head-to-head comparison with open-weight competitors of similar or larger scale (e.g., Llama, Qwen) is provided. The 'sovereign AI' narrative may be persuasive for Japanese procurement processes. However, without transparent and reproducible evaluation, this remains merely a claim rather than a proven advantage.