Heading to ICML 2026 in Seoul next week with @romfbr31 to present Hibiki-Zero🇫🇷🇬🇧🇵🇹🇪🇸🇩🇪[kyutai.org/blog/2026-02-12-h…], Kyutai's latest real-time speech translation model. I'll be giving an oral presentation on July 8 at 10:30 AM KST. Feel free to join if you'd like to learn more!💬

Happening in 15 min at the Scaleway booth, 3rd floor. Come and chat with us !

I'll present our recent and on-going work at @kyutai_labs on multimodal generative AI and world-models tomorrow, Thurs. 18th of June at the Scaleway booth at Vivatech, Hall 7.3 | Booth 3F10, at 10:30am. Come and see us! Also come see @GradiumAI at the AWS booth.

Hypnotizing to watch. Great work, co-authored by our very own @nico_dufour

We upgraded Gradium TTS for the cases voice agents can't get wrong: phone numbers, codes, email addresses read back right the first time. Couple of examples: English: 97% on emails, top of the field. French: leads every competitor we benchmarked. Samples methodology → gradium.ai/blog/gradium-tts-…

Making full-duplex speech dialog models more interactive, reactive, engaging -- and human-y 🤖🗣️ An incredible effort driven by @atsumoto_ohashi, in collaboration with @neilzegh, @honualx, and myself.

Moshi is great but sometimes not fully exploiting its full-duplex abilities. @atsumoto_ohashi applied carefully crafted RL rewards on Moshi's output given some real inputs to improve interactivity on all axis. Works great on Moshi and the derived PersonaPlex by @nvidia.

Pushing the performance of our on-device TTS model, which supports arbitrary voices.

Today @kyutai_labs and @ELLISInst_Tue launch @kesai_labs! Robot learning is bottlenecked by the cost of physical interaction. Our mission is to advance the efficiency frontier of robust & safe physical AI through fully open and reproducible research. kesai.eu/blog/2026-05-20-kes…

Research and innovation in AI must be at a European scale. @kyutai_labs and @ELLISInst_Tue are partnering to launch @kesai_labs, a new lab for physical AI and autonomous driving. The founding team are rockstars, ready to push scientific excellence, openness and real-world impact.

🙏

This is cool but @kyutai_labs demonstrated this like a year 1/2 ago ago. On an iPhone.

TM’ model seems quite exciting. Glad to see @kyutai_labs works Moshi (from 2024) and MoshiRAG cited as inspirations, ideas and research do flow westward over the Atlantic, and we’ll keep pushing new ones. To be tested!

Three of our papers got accepted at ICML and one at CVPR this year 🎉 We will have researchers on-site for both conferences, so come talk to us if you want to learn more about Kyutai! 👁️ MoshiVis (CVPR’26) → Vision Speech Models: A data- and training- efficient pipeline for omni models built on top of Moshi 🧠 MoshiRAG (ICML’26) → Making speech-to-speech models smarter with the power of RAG and minimal latency 🗣️Hibiki-Zero (ICML’26) → Streaming speech-to-speech translation without aligned data leveraging GRPO ⌛ Kairos (ICML’26) → Recency bias is real, even for LLMs. More details in a future post! #ICML2026 #CVPR2026

Digging the story behind Chernobyl with MoshiRAG. 2026 is the 40th anniversary of the reactor meltdown, a tragic event, unavoidable given the systemic issues in the USSR nuclear program. I have a thought for the firefighters and liquidators that risked or gave their life there.

This work was done by Chung-Ming Chien from @TTIC_Connect while interning @kyutai_labs last summer. Read the paper at arxiv.org/abs/2604.12928. Interestingly, very similar work was developed by @SakanaAILabs around Moshi, showing this kind of hybrid approach is the way forward!

Hybrid speech-to-speech models with access to external search or text LLMs are the best way to combine low latency full-duplex with the highest capabilities. In Moshi-RAG, Moshi can decide when to "call a friend" and receives a compressed document to guide it.

Step 1: nailing our accuracy on the standard multilingual TTS eval set (from MiniMax) ✅ Step 2: realizing this benchmark does not capture real world use cases, so we're building a harder one⏱️

We are looking for talent able to help us shape @GradiumAI data acquisition, annotation and QA process, with immediate impact on our line of products and future research. Help us bring speech interaction to the next level. 👉 gradium.homerun.co

If you are doing a PhD on generative AI, model alignment, or speech, and want to get at the forefront of speech modeling research, we are opening PhD level internship at @GradiumAI. Come and join the team behind Mimi, Moshi, Hibiki, and PocketTTS. 👉 gradium.homerun.co/