Sarashina3 mini

Sarashina3 mini is a Japanese-focused language model that SB Intuitions, a SoftBank subsidiary, built from scratch, and it is the main chat model in a family that also includes a smaller version, a safety filter and search models. It was trained mostly on Japanese and English text, then refined by transferring knowledge from a larger model and through several rounds of reinforcement learning. The company says it beats its predecessor, Sarashina2 mini, on every evaluation it ran, roughly matches general-purpose models from OpenAI overall, and pulls ahead on a test of knowledge about Japan. It does not think longer before answering, and the company admits it is weaker than such models at carrying out multi-step tasks on its own. It is sold only to businesses through SoftBank's domestic cloud, and its price, size and how much text it can read at once are not disclosed.