Tanuki

Japanese LLMs pretrained from scratch by the Matsuo-Iwasawa Lab under the GENIAC national program, built by a publicly recruited volunteer team (students, researchers, industry). Tanuki-8B trained on ~1.3T tokens; Tanuki-8x8B is a from-scratch upcycled MoE (~47B total / ~13B active, ~1.7T tokens) that matched GPT-3.5-Turbo on Japanese MT-Bench — at release a leading domestically-trained Japanese LLM. Apache-2.0, with DPO chat tuning and vision variants (Tanuki-8B-vision, 8x8B-vision-exp) built via data synthesis. Direct successor to Weblab-10B.