SB Intuitions Introduces the "Sarashina3 Series" on "Cloud PF Type A"

SB Intuitions Corp. (“SB Intuitions”) has started providing the “Sarashina3 Series,” the latest generation of its homegrown generative AI model “Sarashina,” through “Cloud PF Type A,” a sovereign cloud service provided by SoftBank Corp., on June 30, 2026.

The Sarashina3 Series is a family of generative AI models that includes an enhanced standard model and embedding model with improved Japanese language and instruction-following capabilities, as well as newly added specialized models designed to address increasingly diverse business needs.

Customers using Cloud PF Type A can access five models in the Sarashina3 Series: the standard model “Sarashina3 mini,” the lightweight model “Sarashina3 nano” for high-speed, cost-efficient processing of routine tasks, the guard model “Sarashina3 guard” for safe AI utilization, the embedding model “Sarashina3 embedding,” and the reranking model “Sarashina3 rerank” for advanced search and RAG systems.

Sarashina3 mini and Sarashina3 embedding deliver improved performance through significant enhancements to both pre-training and post-training methodologies.

Key Models Available on Cloud PF Type A

Sarashina3 mini / Sarashina3 nano

These models were pre-trained on more than 30 trillion tokens of large-scale, high-quality data, primarily in Japanese and English. Post-training incorporated state-of-the-art distillation techniques and multi-stage reinforcement learning to enhance capabilities critical to real-world business applications.

Sarashina3 mini is a balanced model that combines performance and speed. It features agent capabilities that can autonomously call external APIs, advanced mathematical reasoning and code generation capabilities, and delivers concise responses suitable for business use cases.

Sarashina3 nano is a speed-focused model that offers exceptional processing speed at low cost. It is particularly well suited for large-scale routine tasks such as data classification and information extraction.

Sarashina3 guard

This model analyzes input text and classifies it as either “Safe” or “Harmful.” Harmful content is further categorized with corresponding category codes, including violent content, information leakage, and assistance with illegal activities. The model can be used as both an input and output filter to support safe AI operations.

Sarashina3 embedding

This model converts text into high-dimensional numerical vectors and is optimized for candidate retrieval from large datasets during the initial stage of high-speed document search. It supports seven task types that can be selected according to specific use cases. In addition, the generated vectors are designed to maintain strong performance even when their dimensionality is reduced.

Sarashina3 rerank

This API analyzes and compares multiple candidate documents against a given query and returns them in order of relevance with high accuracy. Combined with Sarashina3 embedding, it helps reduce hallucinations in RAG systems and improve response accuracy. The API also supports Japanese-language instructions and offers five task types.

For more information about the Sarashina3 Series, please visit:
SB Intuitions Tech Blog: Collection of Articles on the Sarashina3 Series