tsuzumi 2

tsuzumi 2 is NTT's second-generation proprietary large language model, launched in late 2025 and built from scratch for Japanese-language processing with multimodal support (text, image, voice). Its defining trait is efficiency: it runs inference on a single GPU while NTT claims world-top Japanese performance among comparable-size models and parity with larger models in business deployments. NTT reinforced finance, medical, and public-sector knowledge for domain fine-tuning, and Tokyo Online University adopted it for an on-premise campus platform on data-sovereignty grounds.

The honest picture: tsuzumi 2 is proprietary and enterprise-licensed — no open weights, no public benchmarks, and NTT's performance claims are internal evaluations. It is a Japanese-first model; English and other-language capability is secondary, and the closed ecosystem means no community fine-tunes or independent verification of the 'matches larger models' claim.