French AI lab Kyutai launches Moshi, a powerful AI model rivaling GPT-4o and Google Astra.

From Analytics India Magazine: 2024-07-04 08:22:14

Kyutai, a French non-profit AI research lab, launches Moshi, a real-time native multimodal foundational AI model rivaling OpenAI’s GPT-4o and Google Astra. Moshi can understand 70 emotions, speak with accents, and handle two audio streams simultaneously, optimized for CUDA, Metal, and CPU backends. Features include real-time interaction with 200ms latency and ability to run on MacBooks.

Developed by a small team, Moshi integrates text and audio training on the Helium 7B model, offering support for 4-bit and 8-bit quantization. Kyutai chief Patrick Pérez believes Moshi could revolutionize human-machine communication. The lab, founded in 2023 with backing from investors like Xavier Niel, plans to release the full model to contribute to open research in AI.

Kyutai’s approach challenges major AI companies like OpenAI, focusing on open releases and ecosystem development. Moshi’s capabilities could enhance France’s influence in the AI sector. Notably, Kyutai’s effort contrasts with companies like OpenAI, which have faced criticism for delaying releases over safety concerns.



Read more at Analytics India Magazine: French AI Lab Kyutai Releases OpenAI GPT-4o Killer ‘Moshi’