UTokyo

The University of Tokyo is Japan's flagship university; its AI center of gravity is the Matsuo-Iwasawa Lab (Graduate School of Engineering) — the country's best-known academic deep-learning group and startup factory (PKSHA, DeepX, ELYZA; spin-outs are tracked separately) — alongside Tatsuya Harada's Machine Intelligence Lab at RCAST (dual-hatted with RIKEN AIP) and Masashi Sugiyama's ML-theory group (Sugiyama directs RIKEN AIP, Japan's national AI center). Yutaka Matsuo chairs the Cabinet Office AI Strategy Council and directs the AI Robot Association (AIRoA).

The from-scratch line: Weblab-10B (2023, 10B GPT-NeoX on ~600B tokens) was the lab's first large open Japanese LLM, and its GENIAC national-program successor Tanuki (2024) scaled to an upcycled 8x8B MoE (~47B total / ~13B active, ~1.7T tokens, Apache-2.0) that matched GPT-3.5-Turbo on Japanese MT-Bench — then a leading domestically-trained LLM. GENIAC weights ship under the weblab-GENIAC org; the lab also released a PII-filtered 10B-token Japanese pretraining corpus.

Harada's Asagi VLM family (synthetic, license-clean Japanese multimodal data) extends to the Med-Asagi medical VLM; the Iwasawa-senior-authored MMLU-ProX is a widely-used 29-language reasoning benchmark; and robotics is ramping through AIRoA: the AIRoA MoMa mobile-manipulation dataset, with a claimed ~100,000-hour robot-motion dataset and a multimodal robot foundation model in progress (watch items, no public checkpoint). Not filed here: Swallow (Institute of Science Tokyo), LLM-jp (NII), and Matsuo-lab spinoff ELYZA.