rinna

りんな / rinna株式会社

Developing Japanese-language foundation-model derivatives, speech synthesis, and virtual-human technology, this private company adapts global open-weight models into tailored enterprise AI applications for the domestic market.

Private company
Profile as of September 14, 2026
Not ranked: no published revenue or market capitalisation
How the ranking works →

rinna translates global open-weight models into Japanese-specific enterprise applications, offering a credible alternative to general-purpose LLMs. However, its heavy reliance on external base models like Qwen and Llama, combined with complete financial opacity and undisclosed leadership, creates significant execution and geopolitical risks.

Key figures

Revenue: Not disclosed
Revenue growth: Not disclosed
Operating margin: Not disclosed
Net income: Not disclosed
R&D spend: Not disclosed
Capital expenditure: Not disclosed
LINE friends connected: More than 8.6 million
Hugging Face model collections 2025-03-23: 10
Virtual-human supported languages: More than 140
Series A valuation 2021-07-12: Estimated US$1 billion

Overview

Established in June 2020 as a spinout from Microsoft Japan's AI character project, rinna operates as an independent private company focused on Japanese-language generative AI. Rather than building massive foundation models from scratch, the company specializes in adapting and continually pretraining global open-weight models for the domestic market. Its product portfolio spans text, speech, image-language, and virtual-human motion generation, providing a broader surface area than a pure large-language-model developer.

The company bridges the gap between raw foundation models and enterprise utility. Through its Tamashiru Enterprise platform and developer APIs, rinna allows businesses to integrate custom language models with their own proprietary data. While it does not supply the underlying compute infrastructure or semiconductor hardware, its position in the software stack makes it a direct participant in Japan's regional AI adoption. Its virtual-human and conversational AI solutions are deployed in customer-facing roles, including educational institutions and public-service assistants, demonstrating practical commercialization of its multimodal research.

The AI angle

rinna captures AI demand directly through enterprise deployments of custom language models, developer APIs, conversational agents, text-to-speech services, and virtual-human video generation. The company's core offering revolves around its Japanese-adapted model families, including the Qwen-derived Nekomata and Bakeneko series, the Llama-based Youko, and the Gemma-based Baku. By continually pretraining these models on Japanese corpora and applying instruction tuning, rinna provides localized alternatives to global general-purpose models.

Its enterprise solutions, such as Tamashiru Enterprise and Tamashiru Custom, combine these tailored models with selected third-party LLMs and customer-specific data pipelines. This allows Japanese businesses to deploy generative AI without managing the underlying model architecture. The company also monetizes multimodal capabilities, offering virtual-human solutions that support more than 140 languages for use in customer service, education, and public-facing digital interfaces.

Because rinna is a software and model-services provider, its revenue is tied to application-level AI adoption rather than infrastructure capital expenditure. It does not manufacture accelerators, servers, or networking equipment, nor does it operate hyperscale cloud data centers. While the company does not disclose its revenue, product-line breakdown, or customer concentration, its public model releases and enterprise case studies confirm active commercialization of its generative-AI stack.

Technology and moat

rinna's primary technological advantage lies in its Japanese-language adaptation techniques and multimodal integration rather than proprietary foundation-model architectures. Because its most prominent model families—Nekomata, Bakeneko, Youko, and Baku—are derived from Alibaba Cloud's Qwen, Meta's Llama, and Google's Gemma, the company does not own the underlying base-model intellectual property.

Instead, its defensibility stems from the difficult-to-replicate components of localization and deployment. These include curated Japanese training corpora, continual-pretraining methodologies, instruction tuning, and specialized model evaluations. Furthermore, rinna differentiates itself through multimodal assets, combining text generation with proprietary speech-language models, text-to-speech synthesis, and virtual-human motion generation.

This application-layer depth creates switching costs for enterprise customers who rely on rinna's specific prompts, retrieval pipelines, and voice assets. However, the moat is constrained by intense competition. Well-capitalized domestic rivals like NTT, NEC, and Preferred Networks, alongside global giants like OpenAI and Anthropic, offer compelling alternatives. Without disclosed proprietary semiconductor technology, exclusive cloud capacity, or massive proprietary compute clusters, rinna remains vulnerable to rapid advancements by larger base-model developers.

Five-pillar assessment

Scale and market position: A private, niche player in the Japanese AI market, lacking the massive capital, compute infrastructure, and revenue scale of its domestic telecommunications and global hyperscaler competitors.
Technology and R&D: Strong capabilities in Japanese-language continual pretraining, speech synthesis, and virtual-human generation, though heavily reliant on adapting external open-weight models rather than building proprietary foundation architectures.
Supply-chain centrality: Deeply dependent on Alibaba Cloud, Meta, and Google for base models, and reliant on undisclosed cloud providers for compute, while serving a fragmented enterprise customer base.
Financial momentum: Continues to release updated model variants and secure enterprise deployments, but complete financial opacity obscures its actual revenue growth, cash burn, and commercial traction.
Governance and quality: Highly opaque private governance with no disclosed financial statements, an unannounced CEO succession following the founder's 2024 departure, and undisclosed ownership structures.

Supply chain and relationships

rinna operates with significant upstream dependencies on global technology providers. The company relies heavily on external base-model developers, utilizing Alibaba Cloud's Qwen architectures for its Nekomata and Bakeneko series, Meta's Llama for Youko, and Google's Gemma for Baku. It also depends entirely on undisclosed third-party cloud infrastructure, GPU suppliers, and data-center operators for model training, fine-tuning, and enterprise inference hosting.

Downstream, rinna serves enterprise and public-sector customers seeking localized AI applications. While the company does not systematically disclose its client roster or revenue concentration, public examples include the Philippine Department of Trade and Industry, which reportedly uses rinna for a public-service assistant, and Colegio de San Juan de Letran-Calamba, which deployed its virtual-human solution. The company also partners with A.I., Inc. for free-dialogue technology research. This ecosystem position makes rinna a critical localization layer for foreign models entering Japan, but leaves it highly exposed to upstream licensing changes and compute-supply bottlenecks.

Customers

Colegio de San Juan de Letran-CalambaUses rinna's virtual-human solution on its website
Department of Trade and Industry PhilippinesUses rinna in Project TRINA public-service assistant

Partners

Alibaba CloudBase models for the Nekomata series
A.I., Inc.Research collaboration on free-dialogue technology

Competitors

NTTInferredDevelops tsuzumi Japanese-language generative AI
NECInferredMarkets cotomi as a Japanese enterprise AI platform
Rakuten GroupInferredDevelops and distributes Japanese LLMs
Preferred NetworksInferredDevelops domestic foundation models (PLaMo)
OpenAIInferredOffers general-purpose and enterprise foundation models
AnthropicInferredProvides enterprise large-language models
GoogleInferredSupplies Gemma base models and competes in AI
MetaInferredSupplies Llama base models and competes in open-weights

Geopolitics and risk

rinna's reliance on foreign base models introduces substantial geopolitical and supply-chain risk. Its flagship Nekomata and Bakeneko series are built on Alibaba Cloud's Qwen models, exposing the company to potential shifts in Chinese technology export policies or U.S. and allied restrictions on Chinese AI ecosystems. Similarly, its use of Meta's Llama and Google's Gemma ties its product roadmap to the licensing terms and release cadences of American technology giants.

The company also faces cross-border operational exposure through its deployments in the Philippines, which involve public-sector and educational institutions. This subjects rinna to local data residency requirements, public procurement rules, and the reputational risks inherent in deploying virtual humans and conversational AI in government contexts. Domestically, rinna must navigate Japan's evolving AI governance landscape, which increasingly scrutinizes model-training data transparency, copyright compliance, and enterprise data security. Without disclosed proprietary compute infrastructure, the company remains vulnerable to global accelerator shortages and cloud-pricing volatility driven by broader geopolitical tensions.

Risk matrix
Risk	Severity	Why it matters
Dependence on foreign base models	High	Changes in upstream licensing or geopolitics can constrain its roadmap.
Cloud and accelerator dependency	High	Relies on undisclosed external compute for training and inference.
Intense model competition	High	Faces strong domestic and global alternatives, limiting pricing power.
Private-company financial opacity	High	Lack of financial disclosure limits assessment of durability.
Japan AI governance and data regulation	Medium	Must navigate emerging privacy, copyright, and training-data rules.
Cross-border data and Philippine public-sector exposure	Medium	Creates local procurement, data residency, and policy risks.
Reputation and safety risk	Medium	Virtual humans and synthetic voice carry impersonation and deepfake risks.
Customer concentration	Medium	A few enterprise projects could account for a material share of revenue.

Governance and ownership

rinna is an independent private company established in June 2020 to assume the chatbot-AI business originally developed by Microsoft Japan. It operates with extreme financial and structural opacity. The company does not publish audited financial statements, revenue figures, or capital expenditure details.

Governance visibility is similarly constrained. Following the announced departure of founder and CEO Zhan "Cliff" Chen in February 2024, the company has not publicly named a successor, current chief executive, or chairman. The ownership structure, cap table, board composition, and the identities of its Series A investors remain undisclosed. There is no public evidence of state ownership or control by a larger corporate group, but the lack of standard corporate reporting limits external assessment of its financial durability and strategic direction.

What to watch

Whether rinna names a chief executive or chairman following Zhan Chen's February 2024 departure.
Whether the company raises a new financing round or reports a valuation after the 2021 Series A.
Whether Tamashiru Enterprise gains named customers or disclosed recurring-revenue measures.
Whether rinna releases proprietary Japanese base models rather than derivatives of Qwen, Llama, and Gemma.
Whether the company identifies cloud, GPU, or hosting partners that clarify its inference capacity.
Changes in Japanese AI-governance guidance affecting model-data transparency and enterprise sales.

Recent developments

March 2025Released qwq-bakeneko-32b and qwen2.5-bakeneko-32b-instruct-v2.
February 2025Released Qwen2.5 Bakeneko 32B and related instruction-tuned variants.
December 2024Ended the Tamashiru Trial, keeping Tamashiru Enterprise available as an LLM solution.
July 2024Publicized use of its virtual-human solution on Colegio de San Juan de Letran-Calamba's website.

Coverage on AsiaAI.FYI

We haven't written about rinna yet. Subscribe below to get it when we do.

Guides that cover rinna

Japan AI Industry Map

About this profile

Compiled with AI-assisted research from company filings, market data, and published reporting as of September 14, 2026, then reviewed by AsiaAI.FYI. Figures marked Estimate are not company-reported. Check primary filings before relying on any number.

Confidence: C. Company product disclosures and partner references support the operating profile, but audited financial and ownership information is absent.

Main sources: Company corporate and product pages; Developer documentation; Public model repository records; Company-related press releases; Public private-company databases.

All 61 sources

Position in the AI stack

Operating in the model-developer and enterprise-application layers, it adapts foundation models and multimodal AI components for Japanese-language deployments.

Legal name: rinna Co., Ltd.
Headquarters: Tokyo, Japan
Founded: 2020
Founders: Zhan Chen / 陳湛
Employees: 37
Ownership: Private company
Business: Japanese-language generative-AI models and enterprise applications
Sector: AI Software & Models
Website: rinna.co.jp