Skip to content
Back to articles
ai7 min

Sarashina: SoftBank Intuitions free beta with 6 API models — how to try it safely

KGA IT
2026-10-137 min
SarashinaSB IntuitionsSoftBankIA generativaAPIRAGTraduçãoPMEJapãoBeta
Ilustração abstrata: seis nós de API conectados e balões de chat em tons índigo e dourado (serviço Sarashina)

If your company in Japan wants to try a domestic generative model without signing a contract or opening a partner ticket, SoftBank Intuitions (SB Intuitions) just opened the door. On 13 October 2026, the SoftBank subsidiary launched the free beta of the "Sarashina service": six Sarashina3 models via API, plus a web chat for dialogue and translation. Login is with a Google account, with no prior application. There are rate limits, and the beta end date is still undecided.

This is a practical guide for SMBs: what the official release says, what you can try now, where to be careful with data — and what we do not invent (endpoints, parameters and rate numbers are not in the press release).

What the Sarashina service (free beta) is

SB Intuitions describes Sarashina as a domestic generative AI line strong in Japanese. The free-beta "Sarashina service" brings together:

  1. An API with six Sarashina3 models, meant for embedding AI into systems and services.
  2. A web chat to try dialogue (including search-assisted chat and writing) and Japanese↔English translation.

The stated goal of the beta is to collect usage and feedback to guide the product. The company also notes that the Sarashina3 series is already available on SoftBank Cloud PF Type A — so this public beta is a parallel trial channel, not the only enterprise path.

Official service URL: https://sarashina.ai.

The six API models (as the press summarizes them)

ModelRole described by SB Intuitions
Sarashina3 miniGenerate text and chat with the AI
Sarashina3 nanoSimple tasks at high volume and speed
Sarashina3 embeddingSearch / RAG over text and FAQs
Sarashina3 rerankImprove search-result precision
Sarashina3 guardJudge input-text safety
Sarashina3 translateNatural Japanese↔English translation

The 13 October release does not list REST endpoints, context sizes, future pricing or tokens-per-minute tables. For integration details, read the documentation on sarashina.ai after login — do not copy third-party examples without checking.

If your team already builds RAG with open embeddings, compare the pipeline (embedding → retrieval → rerank → generation) with what you use today, for example EmbeddingGemma 2. Sarashina's angle here is having embedding, rerank, guard and generation in one domestic family.

Web chat: what you can try without code

According to the announcement, the public chat covers:

  • Dialogue with the AI, including web search and writing scenarios.
  • Translation Japanese→English and English→Japanese.

That is useful to:

  • validate tone and quality in business Japanese (email, FAQs, drafts);
  • check whether JA/EN translation fits a bilingual team;
  • show non-technical stakeholders what an API PoC might look like.

It does not replace load testing, SLA review or compliance — it is a product experience, not homologation.

How to get in (and what the release says about limits)

  • Google social login, no application form.
  • Free during the beta.
  • There is a rate limit (cap on requests per time window). The press does not publish the numbers — if a script hits the ceiling, treat that as expected and read the docs/ToS.
  • Beta end date: undecided. Do not plan production dependency on the free beta alone.

Usage data may train the model — treat it as a test environment

The press footnote matters most for SMBs:

In the free beta, SB Intuitions may use usage data to improve models and the service. Details are in the service page terms of use and privacy policy.

In practice:

  • Do not send customer personal data, medical notes, cards, named contracts or trade secrets in the chat or beta API.
  • Use synthetic or public text for PoCs.
  • If you fall under Japan's APPI, align the trial with what we cover in APPI for small businesses: legal basis, minimization and a record of what left the company.
  • Anyone who lived through Japan's breach wave alerts knows the risk is not only "the model hallucinated" — it is where the information went.

When (and if) a paid/contract offer with enterprise data handling appears, that is when DPA, processing region and training opt-out matter. This beta announcement does not spell that out.

How to evaluate safely (SMB checklist)

  1. Define 5–10 real tasks (internal JA FAQ, email summary, catalog translation, ticket classification) with fictional data.
  2. Try the chat first: quality, hallucination, corporate tone, JA/EN.
  3. Build a minimal API PoC only after reading docs on sarashina.ai: a documented chat/completions-style call, an embedding call, a translate call — without inventing paths.
  4. Measure what matters to you: perceived latency, JA quality, free cost vs engineering time, fit with your stack.
  5. Compare with what you already pay (OpenAI, Anthropic, Google, etc.) on quality and data risk, not only "it is free".
  6. Record the decision: if it is only an experiment, note the date, Google accounts used and that no customer PII was sent.
  7. Do not bind production to the beta: feature flags, separate keys, a plan B if the beta ends or the rate limit tightens.

For the broader SMB automation context in Japan, see AI automation for small businesses. In Tokyo research, Sakana AI's MASS is another innovation pole — useful so "AI in Japan" is not treated as a single product.

What we still do not know (and will not invent)

  • Exact rate-limit numbers.
  • API endpoints, schemas and parameters.
  • Price or SLA of a future paid tier.
  • Whether/when the beta ends.
  • Definitive training policy for enterprise data on a commercial offer.

Those should come from documentation and terms on sarashina.ai and SB Intuitions releases — not from rumors.

Conclusion

The free Sarashina service beta is a rare window: a domestic Japanese API + chat, Google login, no application friction, launched on 13 October 2026 by SB Intuitions. For SMBs, the value is evaluating JA quality, RAG and translation with synthetic data — and not confusing a free beta with a safe home for customer data. Start in chat, read ToS/privacy, then run a short PoC. If it fits, commercial talks and Cloud PF Type A enter the radar; if not, you spent a few hours and no sensitive data.

Sources


Want this for your business?

KGA IT helps small and medium-sized businesses set up AI assistants, automation and websites. The consultation is free and we reply within 24 hours. Plans start at ¥5,000/month.

Let's solve your technical challenges together.

KGA IT Solutions delivers AI, cloud, and DevOps expertise to address your specific challenges.

Contact Us