What Kyutai does

Kyutai is a Paris-based nonprofit AI research laboratory founded in November 2023 with a mission to build and democratize artificial general intelligence through open science. The organization is funded by founding donors Iliad Group (telecommunications), CMA CGM Group (shipping and logistics), and Schmidt Sciences, and operates with approximately 20 team members comprising researchers, PhD students, postdoctoral researchers, and technical staff. All models, code, and research outputs are released publicly under open-source licenses (MIT, Apache 2.0), with no commercial pricing for its technology.

Kyutai's product portfolio centers on speech-native AI and multimodal systems. The flagship Moshi is a speech-text foundation model that processes audio directly without converting to text, achieving real-time full-duplex dialogue with theoretical 160ms latency. Supporting products include the Mimi neural audio codec (12.5Hz, 1.1kbps, jointly modeling semantic and acoustic information), Kyutai STT (streaming speech-to-text in 1B and 2.6B parameter versions), Kyutai TTS 1.6B and Pocket TTS (a 100M-parameter CPU-real-time model with voice cloning), Unmute (a modular voice AI system enabling any text LLM to listen and speak), Hibiki and Hibiki-Zero (simultaneous speech-to-speech translation with voice transfer), MoshiVis and CASA (vision-language extensions via cross-attention), and Helium-1 (a 2B-parameter multilingual LLM for edge and mobile deployment).

The organization follows a community-led go-to-market strategy through open-source distribution on GitHub and Hugging Face, supported by online demos, technical blog posts, and active social channels. There is no direct customer revenue model; downstream commercial value is captured through Gradium, a Paris-based startup spun out of Kyutai in December 2025 with $70M in seed funding, and through sovereign deployments such as the French Government's DINUM using Kyutai's subtitling technology in the Visio video conferencing platform deployed to 200,000 public agents.

Name

Kyutai

Legal name

Kyutai

Website

https://kyutai.org

Company type

Private

Founded year

2023

Operating status

Operating

Headcount range

11–50 employees

Short description

Kyutai is a Paris-based nonprofit AI research lab founded in 2023 that builds and open-sources speech-native, multimodal, and edge AI models (Moshi, Mimi, Helium, Hibiki, Pocket TTS), funded by Iliad, CMA CGM, and Schmidt Sciences and serving AI researchers, developers, and sovereign infrastructure deployments globally.

Ownership category

akta.pro rank

Product category

Speech AI Software

NAICS

Computer Systems Design and Related Services (5415)

SIC

Services-Prepackaged Software (7372)

akta.pro primary industry

Conversational AI Platforms (chat/voice bots, orchestration) (HDAAAFAC)

akta.pro secondary industries

Speech Translation & Multilingual Speech Tech (HDAAAFAE), Enterprise Foundation Model Integration & APIs (Connectors, Governance, Deployment) (HDAAACAO)

Keywords

• Open-source AI models
• Speech AI research
• Text-to-speech models
• Speech-to-text models
• Multimodal AI systems

Headquarters

HQ city

Paris

HQ country

France

HQ region

Europe

Offices1 record

Markets served

GTM type

B2B

Offering type

Software

Cost components

Personnel, Technology or R&D, Infrastructure, Operations

Revenue model

1. Research Funding and Grants: Kyutai is a non-profit research lab funded by Iliad Group, CMA CGM Group, and Schmidt Sciences. They operate as an open-science organization without a commercial revenue model. All models and research are released publicly under open-source licenses.

Go-to-market motion1 record

Distribution channels3 records

Marketing channels7 records

Core offering

Kyutai is a Paris-based non-profit AI research lab that develops and publicly releases open-source speech and multimodal AI models. Its core offerings include speech-native dialogue systems (Moshi), streaming text-to-speech (Kyutai TTS 1.6B, Pocket TTS), streaming speech-to-text (Kyutai STT), modular voice AI (Unmute), simultaneous speech-to-speech translation (Hibiki, Hibiki-Zero), and vision-language models (MoshiVis, CASA). All models, code, and research are released under open-source licenses for free commercial and research use.

Product overview

Kyutai is a Paris-based nonprofit AI research lab dedicated to open science, building and democratizing artificial general intelligence. The product portfolio centers on speech-native AI technology, with Moshi as the flagship speech-text foundation model enabling real-time full-duplex dialogue. Supporting products include Kyutai STT (streaming speech-to-text) and Kyutai TTS 1.6B/Pocket TTS (text-to-speech models), unified through the Unmute modular voice AI system. Vision capabilities are added via MoshiVis using cross-attention adaptation, while Hibiki/Hibiki-Zero provide simultaneous speech-to-speech translation. The Mimi neural audio codec serves as foundational infrastructure across all audio products. Language model capabilities are provided by Helium-1 (2B parameter multilingual LLM). All products are released as open-source code and model weights for reproducibility.

Differentiator

Problem solved

Functional benefit

Products and services

• Moshi The first speech-native dialogue system that processes speech directly rather than converting to text and back, enabling minimal latency and understanding of emotions and non-verbal aspects of communication. Uses multi-stream architecture modeling user and model audio in parallel.
• Hibiki Simultaneous speech-to-speech translation model for French-to-English translation. Transfers speaker's voice and flow with quality closest to human interpreters. 2B parameter model with 1B mobile version for on-device inference.
• Hibiki-Zero Simultaneous speech-to-speech translation model supporting French, Spanish, Portuguese, and German to English. Uses reinforcement learning eliminating need for word-level alignment data. 3B parameter decoder-only model.
• Unmute Modular voice AI system that enables any LLM to listen and speak using Kyutai's speech-to-text and text-to-speech models. Response latency below one second. Supports function calling and external tool integration.
• Helium-1 2B parameter modular and multilingual language model supporting 6 languages (English, French, German, Italian, Portuguese, Spanish). Designed for edge and mobile deployment with focus on latency and privacy.
• Pocket TTS 100M parameter text-to-speech model with voice cloning capability. Small enough to run in real-time on CPU. Achieves lowest word error rate (1.84) among comparable models while being the only one faster than real-time on CPU.
• Kyutai TTS 1.6B Streaming text-to-speech model with 1.6B parameters based on delayed streams modeling. Used in Unmute for low-latency applications such as voice assistants. Supports multiple voices across US, UK, France, and Ireland.
• Kyutai STT Streaming speech-to-text model optimized for real-time usage with semantic voice activity detection. Available in 1B (English/French) and 2.6B (English-only) versions. Supports PyTorch, Rust, and MLX implementations.
• MoshiVis Vision Speech Model extending Moshi to discuss images while preserving real-time latency and natural conversation abilities. Uses lightweight cross-attention adaptation modules trained on frozen Moshi backbone.
• Mimi Neural Audio Codec Streaming neural audio codec that efficiently models both semantic and acoustic information at 12.5Hz and 1.1kbps while achieving real-time latency. Jointly models semantic and acoustic information using distillation from WavLM.
• CASA Cross-Attention over Self-Attention framework for efficient vision-language fusion. Allows adapting pretrained token-insertion VLMs to enjoy practical benefits of cross-attention with near-constant memory cost and inference cost over video streams.

Quantifiable outcome

• First real-time full-duplex spoken large language model with 160ms theoretical latency (200ms in practice)
• +2 more outcomes

Named customers1 record

Segments3 records

Ideal customer profiles3 records

Technology focussed Yes

API detail

Has API

No

API docs

API detail

Core technology

AI maturity

App detail

AI capability10 records

Feature9 records

Partnerships

Two partnerships are on record, tiered flagship and core.

• GradiumflagshipStrategic or Co-development PartnerGradium was spun out from Kyutai to commercialize ultra-low latency multilingual voice AI technology. Kyutai's research forms the foundation for Gradium's products. Gradium raised $70M in seed funding from FirstMark Capital, Eurazeo, DST Global Partners, and others.
• French Government (DINUM)coreStrategic or Co-development PartnerThe French government's Interministerial Digital Directorate (DINUM) uses Kyutai's subtitling technology in their sovereign Visio platform, deployed to 200,000 public agents. This represents production deployment of Kyutai's research in national AI infrastructure.

Scale indicators4 records

Recent moves7 records

Expansion highlights6 records

Direct peers

• Sesame: Sesame builds a real-time conversational voice assistant (CSM) and explicitly uses Kyutai's Mimi neural audio codec. It directly competes in full-duplex spoken dialogue AI, the same core product category as Moshi.
• Cartesia: Cartesia develops real-time, low-latency voice AI models (state-space models) for developers and enterprises, competing directly with Kyutai's Unmute and Kyutai TTS offerings in the streaming voice-AI stack.
• Hume AI: Hume AI builds an emotionally aware voice and conversational AI platform (EVI) targeting natural, full-duplex spoken interaction — overlapping closely with Kyutai's Moshi and emotion-aware speech-native positioning.
• PlayHT: PlayHT provides real-time TTS and voice cloning APIs targeting developers and enterprises, directly comparable to Kyutai's Pocket TTS and Kyutai TTS products on the same core capability set.

Broad incumbents

• OpenAI: OpenAI ships GPT-4o with real-time voice mode, the most prominent commercial full-duplex spoken LLM. It is a broad incumbent defining the category Kyutai's Moshi targets, with vastly larger distribution and capital.
• Google DeepMind: Google DeepMind develops Gemini with voice capabilities and publishes foundational speech research, competing with Kyutai on both end products and underlying open speech-translation/speech-LLM research.
• Meta FAIR: Meta FAIR released SeamlessM4T and other open speech-to-speech/translation models, the closest open-source incumbent to Kyutai's Hibiki/Hibiki-Zero simultaneous translation research and a frequent benchmark reference.
• ElevenLabs: ElevenLabs is a leading commercial voice-AI platform for TTS, voice cloning, and conversational agents — broadly overlapping with Kyutai's Pocket TTS, Kyutai TTS 1.6B, and Unmute offerings at production scale.

Emerging players

• Mistral AI: Mistral AI is a Paris-based open-weight foundation model lab — the closest European geographic and open-science peer to Kyutai, and a potential partner/competitor for sovereign AI and open-LLM efforts.
• Rime: Rime builds production-grade TTS and voice-cloning APIs for enterprise, addressing the same developer and enterprise voice-AI customer base that Unmute and Kyutai TTS serve.

Market position

Strengths4 records

Weaknesses4 records

Competitive moat5 records

Key risks5 records

Key highlights6 records

Customer concentration

• Xx.com/highlights
• GitHubgithub.com/kyutai-labs
• LinkedInlinkedin.com/company/kyutai-labs

Revenue estimate

Valuation estimate

Number of profiles

Profiles9 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 Kyutai