SB Intuitions (SoftBank AI)

Japan's sovereign AI lab: SoftBank's LLM arm building Sarashina

LLMs & Chat Freemium Has API Open Source
Researched · Published · Reviewed
RECATOOLS Score
7.2 / 10
Capability
7.5
Value for money
6.5
Ease of use
5.5
ASEAN readiness
4.5
API quality
6
Founded
2023
HQ
Minato-ku, Tokyo, Japan
Users
Enterprise-only API (Sarashina); open-weight models publicly available on Hugging Face
Launched
August 2023 (SB Intuitions established)
Developer
SoftBank Corp.

Overview

SB Intuitions is SoftBank's wholly owned LLM subsidiary, founded in Tokyo in 2023 to build Japanese-native models — the open-weight Sarashina2 family (7B-70B, MIT) and a proprietary 460B flagship, now sold commercially as the Sarashina API (Nov 2025).

Advertisement

Pricing

Pricing shown for reference only. These figures reflect RECATOOLS research as of 11 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.

Sarashina2 (Open Weight)
Free
Self-hosted, MIT-licensed, 7B-70B variants
  • MIT license
  • 7B-70B parameter sizes
  • Hugging Face download
  • Self-host only
Sarashina API (Enterprise)
Custom
Hosted Sarashina mini via chat + embeddings API
  • Chat Completion API
  • Embeddings API
  • Japan-first support
  • Quote-based pricing

Use cases

Japanese-language enterprise chatbots and document automation using the Sarashina API Industrial robot control and factory-floor optimisation via Physical AI foundation models (target: 2030) Telecommunications network design and operations using SoftBank's Large Telecom Model built on Sarashina Japanese-language document summarisation, report generation, and meeting-minutes automation AI safety and LLM guardrail research in collaboration with Japan's national research institutes

What you can produce with SB Intuitions (SoftBank AI)

  • Open-weight Sarashina2 LLMs (7B-70B, MIT licence) for self-hosted Japanese-language inference
  • Enterprise Sarashina API with chat completion and embeddings endpoints
  • SoftBank Large Telecom Model: Japan-developed AI for telecom network management
  • Sarashina2.2-TTS: Japanese-centric text-to-speech model
  • Sarashina2.2-OCR: 3B-parameter Japanese and English document OCR model
  • LLM safety and alignment research outputs via NICT collaboration (from February 2026)
  • Participation pathway in the Japan AI Foundation Model Development JV (SoftBank/NEC/Honda/Sony) for industrial partners
Advertisement

ASEAN Perspective

SB Intuitions (SoftBank AI) in Southeast Asia

SB Intuitions' models are currently optimised for Japanese language and Japanese enterprise deployments, with no publicly confirmed ASEAN-language support or regional partnerships as of mid-2026. However, the Japan AI Foundation Model Development consortium's Physical AI ambitions are directly relevant to ASEAN manufacturing: Japan supplies roughly 70% of the world's industrial robots, and factories across Thailand, Vietnam, Indonesia, and Malaysia rely on Japanese automation equipment that this initiative aims to make AI-native. ASEAN governments tracking data-sovereignty approaches to AI may find the Japan model of public-private co-investment and onshore data processing instructive. Any practical ASEAN deployment of Sarashina or Physical AI products will depend on future regional partnerships not yet announced.

RECATOOLS Verdict

SB Intuitions is arguably the most strategically significant AI lab in Japan, backed by SoftBank's compute and a real government funding pipeline. The Sarashina2 family (7B-70B, MIT-licensed) is a legitimate benchmark for Japanese-language performance, and the enterprise Sarashina API — live since November 2025 — actually serves 'Sarashina mini,' a lighter model derived from the 460B flagship rather than the flagship itself. SoftBank also co-founded Japan AI Foundation Model Development in April 2026 alongside NEC, Sony, and Honda, targeting a trillion-parameter Physical AI model for factory and robotics control; AI startup Preferred Networks is a technical partner on that project, not a founding equity member, and practical deployment isn't expected before 2030. Enterprise API pricing is unpublished, and the flagship 460B model's benchmarks can't be independently verified. For Japanese-language work today, the open-weight Sarashina2 releases are the more accessible option.

Independent AI-assisted assessment by RECATOOLS.

What people say

The 460-billion-parameter model that headlines SB Intuitions' pitch isn't actually what most enterprise customers get. When SoftBank launched the commercial Sarashina API in November 2025, it shipped 'Sarashina mini' — a lighter model built using lessons from the 460B flagship, not the flagship itself. SoftBank ran the mini model through internal trials with roughly 20,000 of its own employees before opening it up, a reasonable way to stress-test a Japanese-language model before charging for it.

The open-weight side is more straightforward and genuinely useful: the Sarashina2 family spans 7B to 70B parameters, ships under MIT license on Hugging Face, and gives teams that need Japanese-first, self-hostable LLMs a real option without an enterprise sales call. Pricing for the hosted API isn't published anywhere, so cost comparisons against OpenAI or Anthropic require a direct conversation with SoftBank.

The bigger story in 2026 is Physical AI. In April, SoftBank, NEC, Sony, and Honda formed a new company — Japan AI Foundation Model Development — to build a trillion-parameter model for robotics and factory automation, backed by up to ¥1 trillion in NEDO funding over five years plus capital from Japan's three megabanks and two steelmakers. AI startup Preferred Networks is contributing engineering talent but isn't one of the four founding, equity-holding members, a distinction some early coverage blurred. A limited research release is expected around 2027, full production more like late 2028, and practical factory and robotics deployment isn't targeted until the 2030s — this is a roadmap, not a shipping product.

Summary of public user & expert reviews, compiled by RECATOOLS.

Notable facts

  • Sarashina is named after a classic 11th-century Japanese diary, reflecting the goal of encoding Japanese cultural depth into its models.
  • SoftBank deployed a 4,000-GPU Nvidia B200 DGX SuperPod cluster specifically to train Japanese LLMs, one of Japan's largest AI compute deployments.
  • The Japan AI Foundation Model Development consortium targets a 1-trillion-parameter model focused on controlling robots, not generating text.
  • Japan controls approximately 70% of global industrial robotics manufacturing, giving the Physical AI consortium an unusually rich proprietary training data source.

Frequently asked questions

Are any Sarashina models free to use?
Yes. SB Intuitions publishes open-weight Sarashina2 models (7B, 13B, 70B) on Hugging Face under the MIT licence, free to download and self-host. The proprietary 460B flagship and Sarashina mini API are enterprise-gated with undisclosed pricing.
What is Physical AI and why is Japan focusing on it?
Physical AI refers to AI systems that directly control robots, factory arms, and logistics in the real world rather than generating text or images. Japan is prioritising it because the country dominates global industrial robotics manufacturing and sees an opportunity to lead the next layer: AI that makes those machines autonomous.
How does the ¥1 trillion government programme work?
Japan's NEDO (New Energy and Industrial Technology Development Organisation) administers approximately ¥1 trillion (~US$6.3 billion) in AI support over five years starting fiscal 2026. The Japan AI Foundation Model Development joint venture is the primary expected recipient, targeting a sovereign trillion-parameter Physical AI model.

About this listing

Researched on
Published on
Last reviewed

This entry was compiled from publicly available data including SB Intuitions (SoftBank AI)'s official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with SB Intuitions (SoftBank AI) unless explicitly stated.

Data accuracy

Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.

For the latest details, please refer to SB Intuitions (SoftBank AI) directly →

Spotted something out of date? Suggest an update →

Advertisement