Kyūtai
PulseAugur coverage of Kyūtai — every cluster mentioning Kyūtai across labs, papers, and developer communities, ranked by signal.
- 2026-10-18 product_launch Kyutai released a 2.4 billion parameter VLA model with a memory mechanism. source
10 day(s) with sentiment data
-
Silencio-2 TTS model released with emotion and sound tags
A new open-weight Text-to-Speech model named Silencio-2 has been released, built upon Kyutai's Pocket TTS. This model, with 18.2 million parameters, can incorporate emotion and sound tags directly within the text to inf…
-
Block unveils Dessa, a speech-only full-duplex AI model
Block's AI research lab has introduced Dessa, a novel full-duplex speech model trained exclusively on human conversation audio. Unlike many current voice agents that rely on text-first architectures, Dessa operates dire…
-
New Window Likelihood Method Improves Full-Duplex Speech Models
Researchers have developed a new method called Window Likelihood (WL) to improve full-duplex speech models' ability to handle interruptions and backchannels. Traditional methods using sequence-level Reinforcement Learni…
-
Kyutai releases CADRE video autoencoders for interactive world models
Kyutai has introduced CADRE, a collection of seven open-source causal diffusion video autoencoders. These models are engineered for interactive world models, aiming to reduce action lag. They achieve impressive performa…
-
Kyutai unveils CADRE video autoencoders for faster interactive world models
Kyutai has introduced CADRE, a new set of video autoencoders designed to reduce latency in interactive world models. These autoencoders aim to improve upon existing models by processing video frames more efficiently, ac…
-
Gradium open-sources Phonon on-device TTS models
Gradium has released Phonon, a family of on-device text-to-speech models available in 31M and 40M parameter sizes. The English-only Phonon 31m model, licensed under Apache 2.0, is a compact 64.5 MB and can generate spee…
-
Hugging Face Transformers v5.21.0 adds new models, including document understanding and TTS capabilities
Hugging Face's Transformers library has released version 5.21.0, introducing several new models and updates. Notable additions include MiniCPM-OCR for document understanding, Flint, a mixture-of-experts language model w…
-
Pocket TTS Small models enable CPU voice cloning
Kyutai has released three smaller versions of their Pocket TTS models, with parameter counts of 1.9M, 6.9M, and 18.2M. These English-language models are capable of voice cloning and are designed to run on a central proc…
-
Kyutai and Tencent release new AI memory and robotics models
Kyutai has released REKI 2.4B VLA, a robotic model featuring 32 memory registers and achieving 90.4% on the LIBERO-Plus benchmark with a latency of 21.8 ms on NVIDIA H100 hardware. Concurrently, Tencent has launched Hy-…
-
AI video studio automates content creation, slashing costs to zero
A developer has created an automated video production studio powered by Claude Code, capable of generating up to 40 videos in ten days. This studio researches topics, scripts content, clones the creator's voice using op…
-
Kyutai releases 2.4B parameter VLA model with 3-frame memory
Kyutai has released a 2.4 billion parameter VLA model designed for robot control. This model incorporates a memory mechanism capable of retaining information from the three preceding frames, enhancing its contextual und…
-
New benchmarks and models advance robotic manipulation capabilities · 8 sources tracked
Researchers have introduced several new benchmarks and models for robotic manipulation, focusing on areas like memory dependence, personalization, and industrial constraints. RMBench, for instance, evaluates memory capa…
-
Gradium launches AI voice generator from text prompts
Gradium, a voice AI company, has launched Voice Design, a new tool that generates synthetic voices from text descriptions. Unlike traditional voice cloning, Voice Design does not require reference audio or speaker conse…
-
Kyutai releases open-weight MuScriptor for multi-instrument music transcription
Kyutai, a French AI lab, has introduced MuScriptor, an open-weight transformer model designed for transcribing multi-instrument music into MIDI format. The model is available in three sizes, with the largest having 1.4 …
-
AI voice startup Gradium raises $100M seed, backed by Nvidia
Paris-based AI voice startup Gradium has secured $100 million in a seed funding round, with Nvidia participating as a new investor. This funding will support the company's expansion into the Bay Area to compete for tale…
-
MuScriptor: Open-weight model transcribes multi-instrument music to MIDI
Researchers have introduced MuScriptor, an open-weight model designed for multi-instrument music transcription. Unlike previous models that struggled with complex mixes or single instruments, MuScriptor can transcribe n…
-
MIRA AI model trained on Rocket League data released with demo
A new AI model called MIRA has been released, designed for multiplayer interactive world modeling and trained on data from the game Rocket League. Developed through a collaboration involving General Intuition, Kyutai, a…
-
CPU TTS benchmark: Pocket TTS shows flat RTF, UTMOS struggles with naturalness
A benchmark comparing four CPU-based text-to-speech (TTS) models—Kokoro, Supertonic, Inflect-Nano, and Pocket TTS—reveals distinct performance characteristics. Pocket TTS, utilizing a streaming language model architectu…
-
Kyutai's Pocket TTS offers CPU-based voice cloning from 5s audio
Kyutai has released Pocket TTS, a ~100M parameter streaming language model that generates audio tokens autoregressively. This model is notable for its ability to perform zero-shot voice cloning from just 5 seconds of au…
-
New open-weight MuScriptor model transcribes multi-instrument music
A new open-weight model named MuScriptor has been released for automatic music transcription, capable of converting audio recordings with multiple instruments into detailed note sequences. Developed through a collaborat…