← All companies

Frontier / Company profile

Kyutai

What Kyutai does

Kyutai is a Paris-based, non-profit open-science AI research lab focused on building and “democratiz[ing]” artificial general intelligence through open research. Its work centers on multimodal foundation models spanning speech, text, and vision, with a strong emphasis on real-time and streaming interaction paradigms and on publishing code, checkpoints, and technical write-ups for the research community. Kyutai describes its mission as developing large multimodal models using multiple input modalities (text, sound, images) and inventing algorithms to improve model capacities, reliability, and efficiency, while sharing results back to the broader ecosystem.

News

Company record

Jul 10, 2026 · Official · KyutaiMuScriptor: Automatic Multi-instrument Transcription

Kyutai released MuScriptor, describing it as an open model for multi-instrument transcription to date, with a dedicated hosted web UI for uploading audio and exporting MIDI.

Jul 06, 2026 · Official · Kyutai (blog index)MIRA World Model

Kyutai published a post introducing MIRA as a real-time multiplayer world model project.

Jul 01, 2026 · Official · KyutaiSurflo: Consistent 3D surfaces from a global state

Kyutai published details of Surflo, describing a global-state approach for consistent 3D surface reconstruction with decoding designed to support arbitrary resolution.

Jun 18, 2026 · Official · KyutaiThe FID Lottery

Kyutai published a post on quantifying hidden randomness in generative-model evaluation, framed around how reported FID can be affected by underlying randomness.

Jun 10, 2026 · Official · KyutaiPost-training speech models for better interactivity

Kyutai published work on post-training full-duplex spoken dialogue models with RL aimed at improving interaction behaviors (e.g., interactivity-related timing/behavior aspects).

May 26, 2026 · Official · KyutaiKairos: Understanding Data Temporality Impact on LLM pre-training

Kyutai published Kairos, describing benchmarking focused on how temporal ordering in pre-training data affects time-sensitive factual knowledge.

May 20, 2026 · Official · Kyutai (blog index)Introducing KE:SAI

Kyutai announced a new open-science lab dedicated to physical AI created together with the ELLIS Institute Tübingen.

May 04, 2026 · Official · KyutaiPocket TTS now supports six languages

Kyutai updated Pocket TTS to support six languages (English, French, German, Spanish, Portuguese, and Italian) and describes an on-device CPU-oriented design.

Source map · 3 recurring channels · 18 references

Still resolving: Newsroom

Funding

Latest disclosed valuationNot disclosed
Tracked capital$430M2 sourced rounds
DateRoundRaisedValuationLead / investorsEvidence
Nov 17, 2023Founding funding / initial capitalization (non-profit setup; donor contributions)$100M
Not attributed
+3 investors
Nov 17, 2023Founding / initial capitalization (private non-profit research lab)$330M

What it builds

Moshi

Moshi is Kyutai’s real-time, full-duplex spoken dialogue system described by Kyutai as a speech-native foundation model intended for natural conversation with minimal latency.

source ↗
Hibiki

Hibiki is Kyutai’s speech-to-speech translation technology, which Kyutai says it made available via the public sharing of inference code and model weights after Moshi.

source ↗
Unmute

Unmute is Kyutai’s real-time “voice wrapper” that lets developers add streaming speech I/O to an existing text LLM, by coupling Kyutai speech-to-text and text-to-speech capabilities.

source ↗
Pocket TTS

Pocket TTS is Kyutai’s compact text-to-speech model (100M parameters) designed to run on CPU in real time, with voice-cloning capability; Kyutai also states it was updated to speak six languages.

source ↗
Kyutai STT (streaming speech-to-text)

Kyutai STT is Kyutai’s streaming speech-to-text model architecture intended to provide a latency–accuracy trade-off for interactive applications, with released model variants on Kyutai’s site.

source ↗
MuScriptor

MuScriptor is Kyutai’s open model for automatic multi-instrument music transcription, hosted at muscriptor.kyutai.org, which outputs MIDI (and additional downloadable formats) from an uploaded audio recording.

source ↗
MIRA (world model)

MIRA is Kyutai’s real-time “multiplayer world model” project that Kyutai says is trained on Rocket League gameplay and can simulate multiplayer games at interactive frame rates.

source ↗

Milestones & partnerships