SB Intuitions, based in Minato Ward, Tokyo, began offering a free beta of the “Sarashina service” on October 13, 2026, giving users a chance to try its Japan-made generative AI, Sarashina. The service offers two features: web-based chat and an API that can call six AI models. Anyone can use it for free by signing in with a Google account, with no advance registration required.
Six tools lined up beyond the chat
The first feature to try is the web chat, which lets users interact with AI and translate text right in their browser. But the service’s real highlight may be its API, which runs behind the scenes. An API is an interface that lets users call AI functions from their own systems or apps.
The service offers six models. “Sarashina3 mini” handles text generation and conversation, while “Sarashina3 nano” performs simple tasks at high speed and scale. “Sarashina3 embedding” is designed for document search and RAG. RAG, or retrieval-augmented generation, is a system in which AI searches sources such as internal documents and uses the information it finds to generate answers. The lineup also includes “rerank,” which improves the accuracy of search results; “guard,” which assesses the safety of input text; and “translate,” which specializes in Japanese-to-English and English-to-Japanese translation.
Rather than entrusting everything to one “smart AI,” the lineup assigns different roles: writing, searching, reordering, monitoring and translating. Think of it like a kitchen stocked with a full set of tools for different tasks, rather than being handed a single all-purpose knife.
Free API access could change how companies “try out” AI
When only a chat feature is available, individuals may simply ask, “What kind of answers does it give?” With the API also open to users, companies can integrate it into their own systems and see whether it suits their intended uses. For teams considering applications such as handling internal inquiries, searching documents or supporting multiple languages, the barrier to evaluation should be lower.
The models’ division of labor also offers insight for organizations considering adoption. For example, a company building an internal document search system could use embedding to prepare documents for easier searching, rerank to improve the order of search results, and mini to generate answers. Adding guard could also enable safety checks on input. This is just one example of a use case suggested by the lineup, but choosing models according to the task at hand is likely to be key to using AI in the years ahead.
Since the Sarashina series has been developed as a large language model (LLM) with strong Japanese-language capabilities, this gives developers and companies a chance to assess its performance in Japanese for themselves. The company also introduces a tokenizer designed for Japanese—a system that breaks text into units AI can process—and says it can represent the same text using half as many tokens as a U.S. model. Fewer tokens can reduce processing demands and costs, making this an important consideration for people using AI in Japanese.
Terms to check before trying it
Although the service is free, there are conditions. The number of requests allowed within a given period is limited, so it is not suited to sending large volumes of data all at once. It is intended for small-scale trials.
The end date for the beta has not been determined. The company also says it may use user data to improve service quality. Anyone testing the service for work would be wise to check their company’s rules before deciding whether to enter confidential or personal information.
SB Intuitions is a SoftBank subsidiary that conducts research and development in generative AI, with a focus on the Sarashina series. The company says it will use usage data and feedback gathered during the beta to inform future product development and feature expansion. By making a homegrown AI available for anyone to try, the launch also opens the door to a public experiment in which user reactions can directly shape future improvements.
Comments