SoftBank unit opens six Japanese Sarashina AI models for free beta use

What it's about

SB Intuitions, a subsidiary of the SoftBank group, launched a free beta of the "Sarashina Service" on October 13, 2026. Sarashina is a family of large language models (LLMs, meaning AI systems that understand and generate text) specialized in Japanese. The service is open to individual users in the browser and to developers through an application programming interface (API). According to the available sources, the model weights are not offered for download. This is an online service.

The six models

The service bundles six models for different tasks:

RAG (retrieval-augmented generation) means a model first searches for matching documents and then uses them as the basis for its answer. Embeddings are numerical representations of text that let you compare how close two texts are in meaning. A rerank model then checks the retrieved candidates once more for how useful they are for the question.

A post on note.com explains the split using a shop as an example. One model sorts customer questions into categories such as "delivery," "product" or "returns." Another writes the answer. Separating the models makes errors easier to track down. This is the author's thought experiment, not a field report. The post also says you need to check how fast and accurate the models actually are.

Access and terms

Sign-in works through a Google account. According to DXマガジン, no prior application is needed. In the browser you can chat with the AI, draft texts, use web search and translate between Japanese and English. Through the API, companies and developers can build the models into their own systems and test whether they hold up for real tasks.

Use is free for both chat and API. There are limits, though: - The number of requests per time unit is capped. - No date has been set for the end of the beta. - Usage data may be used to improve the quality of the services and AI models.

SB Intuitions plans to analyze usage data and feedback and feed them into further product development. According to DXマガジン, the background is that companies are increasingly moving beyond testing generative AI and building it into business systems. They combine text generation with search, translation and safety checks.

The note.com post advises starting with made-up text or public information only. If you want to process customer data or unpublished plans, read the terms of use and privacy policy first. The guard model only judges whether text is harmful. It doesn't change the data retention rules, and it doesn't guarantee correct answers.

Background: Sarashina at SoftBank

According to note.com (citing SB Intuitions), the Sarashina3 models have been offered through "Cloud PF Type A" since June 30, 2026. So what's new is mainly the easy-to-reach service for newcomers.

Light Reading reported earlier on the predecessor. Alongside its half-year results to September 30, 2025, SoftBank announced the commercial launch of Sarashina mini for November 28, 2025. That model is a compact variant built on the experience with the in-house 460-billion-parameter LLM "Sarashina." Parameters are, roughly speaking, the dials a model learns during training. The API then offered two functions: chat completion for text generation and embeddings for semantic relationships. Companies were meant to use it for tasks such as proofreading and summarizing documents, drafting reports and minutes, coding help and building multi-agent systems. About 20,000 SoftBank employees tested the model in advance, starting in June 2025. SoftBank and OpenAI also founded SB OAI Japan, a 50:50 joint venture. It isn't directly related to the current beta.

What you need to try it

All you need is a browser and a Google account. You don't need your own graphics card. The models run on the provider's servers. The service is aimed at Japanese, so the interface is probably in Japanese. Mac, Windows, Linux and smartphones all work without trouble.

Takeaway

For developers and companies in Japan, this is a low-barrier way to compare several matching models for search, translation and moderation. The pipeline design (search, re-rank, answer, check) fits the enterprise RAG trend. What's unclear is how the models stack up against international alternatives. The sources give no benchmarks, and the note.com post only points to the vendor's own ratings. The future is unclear too: it's not settled whether or for how long the service stays free. If you want to use business data, clear up how usage data is handled first. For readers outside Japan, the news is more interesting than urgent, since the models can't be downloaded openly.