On October 13, 2026, SB Intuitions launched a free beta version service of its domestically developed generative AI, “Sarashina.” This move is driven by the rapidly growing demand from developers and enterprises to pre-test AI technology in a domestic environment with high data security.

Overview of the Free Beta Version and System Requirements

SoftBank subsidiary SB Intuitions Corp. announced the launch of a free beta version of “Sarashina Service,” which allows users to experience the domestic generative AI “Sarashina,” starting October 13, 2026. This service targets not only developers and IT engineers but also individual users who wish to try out generative AI, and it is available without prior registration. Users can access the service by logging in via social login using a Google account on a dedicated web portal site.

The features provided are of two types: an API function that allows you to verify integration into systems and applications, and a chat function that enables conversation and translation in the browser. While there is a rate limit restricting the number of requests per given time period, the service is free to use. As corporate AI adoption shifts from mere experimentation to the stage of integration into actual operations, expectations are growing for an environment that can be immediately used for actual development and verification. Please refer to the figure below.

6 Specialized AI Models Released and Their Roles

In the free beta version, a total of six AI models optimized for different use cases have been released. Users can combine these models via the API to suit their specific needs.

• Sarashina3 mini: A standard model for text generation, AI dialogue, and complex tasks
• Sarashina3 nano: A Lightweight Model for High-Speed, Low-Cost, and Large-Scale Processing of Routine Tasks Such as Information Extraction and Data Classification
• Sarashina3 embedding: An embedding model that converts text into numerical vectors and serves as the foundation for Retrieval-Augmented Generation (RAG)
• Sarashina3 rerank: A reranking model that calculates the relevance scores of search results to improve RAG answer accuracy
• Sarashina3 guard: A safety model for high-speed detection of inappropriate expressions and harmful content in input text and output sentences.
• Sarashina3 translate: A translation-specific model that enables natural bidirectional translation between Japanese and English.

By dividing the processes that were previously handled by a single chat AI into specialized models for generation, lightweight processing, safety evaluation, search, and translation, it is possible to significantly improve the overall processing efficiency and safety of the system.

Sarashina3’s Technical Features and Data Sovereignty

30 Trillion-Token Training and Proprietary Technical Architecture

The Sarashina3 series is not a fine-tuned overseas model, but rather a fully scratch-built domestic large language model independently constructed starting from the pre-training stage. Pre-training was conducted using a massive amount of high-quality data—approximately 30 trillion tokens, centered on Japanese and English. To ensure data quality, curriculum learning, which involves training by gradually increasing difficulty, and strict quality filtering have been adopted.

In the post-training process, we introduce OPSD (Online Policy Self-Distillation) technology—which allows the model to relearn from its own outputs without depending on the generation results of external models—as well as multi-stage reinforcement learning. This enables independent evolution free from reliance on external technologies. Furthermore, to properly evaluate the politeness, nuances, and conciseness specific to the Japanese language, evaluations are conducted by a dedicated human team and fed back into the reinforcement learning process. In embedding models for RAG, MRL (Matryoshka Representation Learning) technology is also employed, maintaining high search accuracy while reducing the number of dimensions by approximately 40 percent compared to conventional models, thereby achieving both lower database storage costs and high-speed searching.

Sovereign Cloud and Ensuring Data Sovereignty

In recent years, interest has been growing regarding data sovereignty concerning the storage locations of data input into generative AI and training data, particularly in fields such as finance, healthcare, and government agencies. Sarashina3 will be provided through “Cloud PF Type A,” a sovereign cloud operated by SoftBank in its domestic data centers.

In this environment, inputted prompts and generation results are never used to retrain the AI models. Furthermore, because all data and access logs are strictly managed and stored in data centers within Japan, you can avoid risks such as information being sent overseas or becoming subject to foreign laws. Additionally, by integrating Sarashina3 guard—a dedicated guardrail model—the system can detect and block risks of data leaks and extreme expressions at the model level. This makes it possible to build a secure AI operational framework that complies with Japanese legal systems and ethical standards, providing powerful infrastructure for enterprises and local governments that demand strict compliance.

Application to Business Implementation and Future Developments

Practical Implementation and RAG Building for Enterprises and Developers

With the public release of the Sarashina service API as a free beta, developers and IT engineers can now proceed with technical verification and prototype development (PoC) for commercial system integration without incurring initial costs. Regarding integration with existing systems, since an OpenAI-compatible API structure has been taken into consideration, it can be easily incorporated with minimal code modifications such as changing the base URL and keys.

Especially when building advanced RAG (Retrieval-Augmented Generation) systems utilizing internal documents, the model integration of the Sarashina3 series truly demonstrates its value. First, Sarashina3 embedding vectorizes a large volume of internal manuals and FAQ data with high precision, and during retrieval, Sarashina3 rerank reorders them in descending order of relevance. Based on that narrowed-down context, Sarashina3 mini generates natural and accurate response text, and finally, Sarashina3 guard determines the safety of the input and output—completing this entire series of processes solely with domestic AI. This makes it possible to achieve both operational efficiency and security assurance. Please refer to the configuration diagram below.

Future Outlook and Key Points to Watch

SB Intuitions plans to collect and analyze actual usage data and feedback through the newly launched free beta version, using these insights to improve model accuracy and expand new features. Regarding the paid commercial platform Cloud PF Type A, provisioning in the East Japan region began ahead of schedule in June 2026, with expansion to the West Japan region scheduled for October 2026. This will also address domestic BCP (business continuity planning) measures and distributed operation needs.

Furthermore, the company initiated a joint research project with the National Institute of Information and Communications Technology (NICT) in February 2026, aimed at achieving both advanced safety and high Japanese language proficiency. By integrating NICT’s abundant Japanese language resources and evaluation infrastructure, the partnership aims to suppress hallucinations and advance alignment technology. Looking ahead, the company also envisions the development of ultra-large-scale models in the 1-trillion-parameter class and multimodal models, and its future technological innovations are drawing attention as a core of sovereign AI that protects Japan’s data sovereignty.

Reference Page

• [Free Beta Version of “Sarashina Service” Now Available, Enabling Access to Domestic Generative AI “Sarashina” | Press Releases | SB Intuitions Corp.] https://www.sbintuitions.co.jp/news/press/20261013_01/
• [Sarashina Service Official Portal] https://sarashina.ai

【#Sarashina #SBIntuitions #DomesticGenerativeAI #DataSovereignty #AIServices】