Kyutai

“Assistive technologies based on open foundation models can improve the life of countless people” With Adopt AI - Grand Palais, the magnificent Grand Palais resonated for two days with the promises of AI. Of particular interest to a non-profit open science lab like Kyutai: how AI is reshaping healthcare. Our CEO, Patrick Pérez, participated to an AI for Health’s panel, with a few strong messages: - A nimble compact research team, focused on well chosen projects, can push frontier IA for a better world; - This is what Kyutai does, for instance in the domain of voice technologies,  which are essential for communication, inclusion and accessibility; - Real-time speech and vision AI models that can run on-device and are fully open are key to the development and the broadest adoption of novel assistive technologies; technologies that can improve the daily life of countless people. Patrick also hinted at an upcoming open-source project designed to support patients who have lost the ability to speak. More on this very soon. This panel was shared with Paul Addison - Medtronic Hugo Ceulemans - Johnson & Johnson Innovative Medicine Hela Ghariani - Ministère de la Santé and moderated by Paul de Balincourt - Artefact  #AdoptAI, #AIforHealth et #AdoptAISummit2025

The frontier of world models and agents is overwhelmingly in Europe. Kyutai is excited to partner with General Intuition to continue pushing that edge through open science!

Launching Kyutai x General Intuition: Research collaboration at the frontier of world models and spatiotemporal reasoning World models and agents capable of spatiotemporal reasoning are the next frontier of AI. As builders of this technology, we have a responsibility to educate. This is impossible without both unique data and world-leading talent. Through our research collaboration, both labs will share data and talent seamlessly. The necessary data lives inside only a few companies, and isn’t available on the internet. With 1B+ ground truth action labeled video uploads per year, General Intuition owns the largest, most diverse, ground truth action labelled dataset. Kyutai utilizes the data for open research, pursuing their culture of publishing on the frontier. Talent moves between the two labs, switching freely between open-ended research and application.  Join us in NYC and Europe - open roles linked in the comments

🎙️ Monde Numérique a posé micro et caméra chez Kyutai. Patrick Pérez et Jerome Colombain ont longuement parlé  de modèles ouverts vocaux, visuels et temps réels ; De communication entre l’humain et la machine, de la façon dont on pourra bientôt interagir avec les contenus audio et vidéo, et converser par dessus la barrière de la langue ; De comment une petite équipe de recherche pointue et audacieuse peut produire des percées majeures. ⬇️ Lien vers le podcast et sa vidéo en commentaire. 

🚨 Wanna join a laser-focused AI research lab in Paris? 🚨 Kyutai is hiring. Kyutai is a daring research lab where a small elite team pushes the boundaries of AI and shares its discoveries with the world. It is all about focus, passion and speed. With no useless dissipation, all the energy is devoted to build and ship impactful models.  Based in central Paris and unique of its kind in Europe, Kyutai is a highly-funded non-profit lab where top-notch research meets the real world, paving the way to tomorrow's technologies. Accelerating our journey, we are seeking new exceptional talents in fields like language and speech models, machine listening, computer vision, generative models and world models. 👋 Application via our website (link in comments ⬇️ ) or DM.

Our open research on portable models that can hear, see and converse, that are efficient, modular and extensible, is paving the way to new technologies.

Try asking your favorite speech LLM whether you’re speaking in a low voice or a high voice: it won’t be able to tell you. But why? These LLMs are trained directly on audio so they should understand speech-specific information like the sound of the voice. It seems natural that by taking the text LLM recipe of “big data + big models = intelligence” and applying it to audio, we should get amazing speech models. In practice, we’re still far from this speech LLM dream. Modeling speech is more difficult than text, for a few reasons. The biggest one is the fact that text is a lot more compressed: a sentence might consist of ten words, but spoken out loud, it consists of tens of thousands of samples. LLMs have trouble staying coherent over that many steps. We’re releasing an in-depth tutorial about neural audio codecs, the de-facto standard way to solve the compression problem for audio. We start from the basics and build up all the way to Mimi, the state-of-the-art codec made at Kyutai. Read the article: https://lnkd.in/dYBFJ6nb

“Open multimodal AI models are changing the way we experience media content and access information” At the Mind Media Day in Paris, our CEO, Patrick Pérez, shared an optimistic message with press and media groups: they possess precious assets, which cutting-edge AI technology will make even more accessible, more engaging, more informative and, ultimately, more valuable. Building on open multimodal models, like those that Kyutai shares, they can easily and safely explore new experiences, new tools, new services. The novel audio experience launched by La Provence, with Kyutai’s text-to-speech (TTS) inside, is a great example: Written articles are turned into lively audios. One day, they will be interactive.       Thanks to mind Media and Jean-Michel De Marchi.

“With world-class open research, let’s make AI a common good for a better world.” At Kyutai, we believe strongly in that mission, and work hard towards it. Our CEO, Patrick Pérez, had the pleasure to discuss this very topic on stage with Anne Bouverot at the Common Good Summit. Organised by Challenges, at the beautiful "Maison de l'Océan" in Paris, the event lined up speakers like Esther Duflo (Economy Nobel laureate), Michael Jordan (immense AI research scientist) and Xavier Niel (one of Kyutai's founding donors with Eric Schmidt and Rodolphe SAADE). Patrick explained how we strive for scientific excellence, impact-driven projects and openness, building and open-sourcing compact generalist models that can hear, see, talk and more.     Thanks to Tom Rouffio and Gilles Fontaine

Unmute turns a text LLM into a voice AI. At unmute.sh, it’s Mistral AI's Mistral-Small-3.2-24B, making it fully open-source. Play a quiz game with a snarky host, catch up on tech news, or just hang out and talk. Or modify it to do anything you want!