What SB Intuitions does

SB Intuitions is a Japanese research and development company founded in October 2023 and headquartered in Tokyo, focused exclusively on building generative AI foundation models specialized for the Japanese language and culture. The company's flagship output is the Sarashina series of large language models, trained from scratch on Japanese-dominant corpora (approximately 1 trillion Japanese tokens for Sarashina1 and 2.1 trillion tokens covering Japanese, English, and code for Sarashina2), with a custom Japanese-optimized tokenizer that delivers materially better tokenization efficiency than multilingual incumbents. The product portfolio now extends across text (Sarashina1/2/2.1/2.2/2.2-Instruct from 0.5B to 70B+ parameters, including the 8x70B MoE model), text embeddings (Sarashina-Embedding-v1-1B and v2-1B), vision-language models (Sarashina2-Vision and Sarashina2.2-Vision-3B), Japanese encoder models (ModernBERT-Ja, 30M-310M), and text-to-speech (Sarashina2.2-TTS). The company also publishes and maintains the JMTEB embedding benchmark and a Safety Boundary Test dataset for Japanese LLMs, positioning it as a measurement authority in the Japanese AI ecosystem.

The company's technical differentiation rests on three pillars: from-scratch Japanese training rather than English-first adaptation, domestic-only data residency (all data kept in Japanese data centers), and an open-distribution posture — most Sarashina models are released under the MIT license on HuggingFace, enabling research and commercial reuse. The business model combines this open-weights strategy with a proprietary 'Sarashina' service platform aimed at Japanese enterprises and research institutions that require domestic data handling and Japanese-language optimization. The company holds ISO/IEC 27001 and JIS Q 27001 ISMS certification covering commercial generative AI operations and is led by a technical bench including CRO Noboru Kuno, CTO and Principal HPC Engineer Toshiaki Hishinuma, and an Information Security Officer, supported by specialist team leads across NLP, Speech, Vision, Responsible AI, and Data Foundation functions.

SB Intuitions is privately held with a strategic relationship to SoftBank — confirmed as parent company in source materials — providing preferential access to HPC infrastructure and enterprise distribution. The company does not publicly disclose revenue, pricing, or funding details, and its commercial customer base is not named in available sources. Operating geographies are explicitly domestic (Japan), with global visibility achieved via open-source model distribution and international academic conference participation (ACL, COLING, INTERSPEECH, SIGDIAL).

Name

SB Intuitions

Legal name

SB Intuitions株式会社

Website

https://sbintuitions.co.jp

Company type

Private

Founded year

2023

Operating status

Operating

Headcount range

101–250 employees

Ownership category

akta.pro rank

Product category

Generative AI Foundation Models

NAICS

Scientific Research and Development Services (5417), Custom Computer Programming Services (541511), Research and Development in the Physical, Engineering, and Life Sciences (54171)

SIC

Services-Computer Programming, Data Processing, Etc. (7370), Services-Computer Programming Services (7371)

akta.pro primary industry

Question Answering & Retrieval-Augmented Generation (RAG) (HDAAADAC)

akta.pro secondary industry

B2B SaaS & Enterprise Personalization (CRM, Sales, Support) (HDAAAGAJ)

Keywords

• Japanese language models
• Large language models
• Generative AI development
• Vision-language models
• Text embedding models

Headquarters

HQ city

Tokyo

HQ country

Japan

HQ region

Asia

Markets served

GTM type

B2B

Offering type

Software

Cost components

Personnel, Technology or R&D, Infrastructure, Operations, Marketing or Sales

Distribution channels2 records

Marketing channels7 records

Core offering

SB Intuitions is a research and development company that develops and releases the Sarashina series of Japanese-specialized generative AI large language models, trained from scratch on Japanese-dominant corpora with all data managed within Japanese data centers. The company offers base LLMs, instruction-tuned variants, vision-language models, text embedding models, text-to-speech models, and encoder models under the MIT license via HuggingFace, with commercial deployment available through the Sarashina Service Platform for Japanese enterprises and research institutions.

Differentiator

Problem solved

Functional benefit

Brands

• Sarashina: Series of Japanese large language models developed by SB Intuitions, including Sarashina1, Sarashina2, and Sarashina3 variants with various parameter sizes from 0.5B to 70B+.
• Sarashina Lab
• Sarashina Tech Blog
• JMTEB

Products and services

• Sarashina Series Large Language Models Series of Japanese-specialized large language models developed from scratch by SB Intuitions on Japanese-dominant corpora. Available in sizes from 0.5B to 70B+ parameters (including MoE architecture), released under MIT license for open research and commercial use via the HuggingFace Model Hub. Targeted at Japanese enterprises and research institutions requiring top-class Japanese language performance, translation, and question answering.
• Sarashina2.2-Instruct Models Instruction-tuned versions of the Sarashina2.2 pretrained models, fine-tuned using supervised instruction tuning and Direct Preference Optimization (DPO). Released in 0.5B, 1B, and 3B parameter sizes under MIT license, achieving the highest benchmark scores in their parameter range on Japanese tasks including Elyza-tasks-100, Japanese MT Bench, and English MT Bench.
• Sarashina-Embedding Models Japanese-specialized text embedding models (v1-1B and v2-1B) developed using two-stage contrastive learning (weakly-supervised followed by supervised) from Sarashina2.1-1B and sarashina2.2-1B base models. Designed for semantic similarity, retrieval, classification, clustering, and reranking tasks, including RAG and knowledge retrieval pipelines in Japanese-language applications.
• Sarashina2-Vision Models Japanese-specialized Vision-Language Models (VLMs) combining image understanding with Japanese language processing. Released in 8B, 14B, and 3B parameter sizes, enabling multimodal AI applications that interpret visual inputs in Japanese-language contexts.
• Sarashina2.2-TTS Japanese text-to-speech model that generates audio from Japanese text input. Developed by SB Intuitions' Speech AI team alongside spoken dialogue systems combining speech recognition, language understanding, and speech synthesis. Released under the Sarashina Model NonCommercial License Agreement.
• ModernBERT-Ja Japanese BERT-style encoder models in 30M, 70M, 130M, and 310M parameter sizes, trained with masked language modeling on approximately 4.4T tokens. Designed for text classification, semantic understanding, and downstream Japanese NLP tasks, achieving highest performance among same-scale encoder models.
• JMTEB Benchmark Japanese Massive Text Embedding Benchmark (JMTEB) developed by SB Intuitions to evaluate Japanese text embedding models. Covers 6 task categories across 16 datasets including clustering, classification, semantic textual similarity, pair classification, retrieval, and reranking. Used as the standard evaluation framework for Japanese embedding model research and development.

Quantifiable outcome

• Sarashina2-70B achieves top Japanese QA performance with 83.36 average score across AI王, abc, JEMHopQA, NIILC, JComQA, JSQuAD datasets
• +3 more outcomes

Segments1 record

Ideal customer profiles1 record

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

AI capability12 records

Feature7 records

Recent moves6 records

Expansion highlights5 records

Direct peers

• Stockmark: Japanese AI company developing Japanese language models and document intelligence products for enterprise customers. Directly comparable in Japanese-NLP focus and enterprise go-to-market in Japan.
• Preferred Networks: Japan-based deep learning company building large-scale models and AI infrastructure (including its PreferredAI/PFNLLM efforts). Directly comparable as a domestic AI lab pursuing Japan-focused foundation model development with industrial partners.
• ELYZA: Japanese AI startup developing Japanese-specialized LLMs (ELYZA-japanese-Llama-2 series and successors) and offering enterprise LLM services. Comparable as a niche Japanese LLM builder targeting the same enterprise and developer customer base.
• Sakana AI: Tokyo-based AI research lab developing Japanese-specialized foundation models with a focus on efficient and evolutionary model training. Closely comparable as a Japanese AI lab competing in the same enterprise and research customer base with similar language-specific ambitions.
• rinna (Microsoft Japan): Japanese conversational AI brand under Microsoft Japan that builds Japanese-tuned LLMs and dialogue models. Comparable as a Japan-specialized LLM developer competing for similar Japanese enterprise and developer mindshare.

Emerging players

• Cohere: Enterprise-focused LLM provider offering multilingual models and embedding/RAG APIs. Comparable as an enterprise LLM and embedding API competitor, though without the same Japanese-language specialization.

Broad incumbents

• Anthropic: Frontier AI lab behind Claude models, widely used by Japanese enterprises via API and increasingly available in Japanese. Comparable as a global LLM incumbent whose models are an alternative procurement choice for Japanese customers.
• OpenAI: Creator of GPT-4/5 and the ChatGPT enterprise platform, the dominant global LLM provider that Japanese enterprises also adopt. Comparable as a frontier general-purpose LLM provider competing for the same Japanese enterprise workload.
• NTT (NTT Human Informatics Laboratories): Large Japanese telecom and R&D group with internal LLM initiatives (e.g., tsuzumi) and significant AI research. Comparable as a major Japanese institution building Japanese language models, though as part of a much broader corporate and telecom portfolio.
• SoftBank Corp. (parent AI activities): Strategic parent and Japanese telecom operator deploying SB Intuitions' models internally and investing heavily in AI infrastructure. Related as the parent organization and overlapping infrastructure partner rather than a clean external competitor.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat5 records

Key risks6 records

Key highlights7 records

Customer concentration

• Xx.com/sbintuitions
• GitHubgithub.com/sbintuitions
• Facebookfacebook.com/sbintuitions
• LinkedInlinkedin.com/company/sbintuitions

Revenue estimate

Valuation estimate

Number of profiles

Profiles5 records

Funding overview

Funding rounds

Investors

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about SB Intuitions