Kyutai open-sources Moshi, a 200ms real-time voice AI that runs locally

Moshi is an open-source, real-time voice AI built by French non-profit research lab Kyutai. It lets you have fluid, natural voice conversations with near-zero lag, clocking an end-to-end latency of just 200 milliseconds, and you can run it entirely on your own machine.

A tiny team of eight researchers built Moshi from scratch in just six months. Instead of chaining separate speech-to-text, LLM, and text-to-speech tools together, it handles audio directly so it can genuinely think while it talks.

What it can do

The standout feature here is simultaneous two-way audio. Moshi can listen and talk at the same time, meaning you can interrupt it mid-sentence or talk over each other just like a normal human conversation.

It handles an impressive emotional range, sporting 70 different speaking styles and emotions along with multiple accents. Kyutai even brought in a professional voice artist to nail the tone and cadence. Alongside the voice stream, Moshi outputs text in real time so you can read along.

Under the hood, it was trained directly on annotated speech rather than plain text, and fine-tuned on 100,000 synthetic dialogues. That direct audio training gives it an ear for inflection and subtext that typical voice bots completely miss.

Hardware specs and open availability

Unlike giant cloud models, Moshi is designed to run locally on consumer-grade hardware like laptops, requiring no internet connection. It comes with backend support for CPU, Nvidia CUDA, and Apple Metal, keeping all your conversations completely private on-device.

Kyutai plans to release Moshi's full code and framework under an exceptionally permissive open-source license. To keep things safe, they are also building built-in audio watermarking and synthetic signature detection tools so people can trace AI-generated audio.

Why it matters

While big tech players keep demoing slick voice assistants behind walled gardens, Kyutai just proved an eight-person lab can build a 200ms real-time voice model and hand it out for free. For developers building customer support bots, interactive NPCs, or local assistants, this is easily the most exciting open weights project to watch.

The real test will be how smart its reasoning feels once the wider open-source community gets to run it locally, but bringing instant, conversational voice to everyday laptops changes the baseline for what open AI can do.