
Kyutai Moshi
Not yet source-checkedOpen-source, full-duplex speech-to-speech foundation model with sub-200ms latency
Machine-drafted from the vendor's site · not yet source-checkedProfile updated Link checked How we rate
In short
Moshi is an open-source, full-duplex speech model by Kyutai offering sub-200ms latency. It runs self-hosted on your own GPU or Apple Silicon with no API costs.
Researchers, voice-AI engineers and indie teams who need a self-hosted, low-latency spoken-dialogue model with open weights and no per-minute API costs.
Product teams that want a managed voice API, multilingual coverage out of the box, or GPT-4o-level knowledge and reasoning wrapped in speech.
Moshi is a speech-text foundation model released by the French non-profit AI lab Kyutai for real-time, full-duplex spoken dialogue. Unlike traditional voice assistants that pipeline ASR to an LLM to TTS, Moshi natively processes two continuous audio streams — the user's microphone and its own synthesized speech — while simultaneously predicting text tokens as an 'inner monologue' that grounds what it says. The result is a conversational voice model that can be interrupted, back-channel ('mm-hmm'), talk over the user, and respond with roughly 200ms of end-to-end latency on an L4 GPU, well below the human turn-taking threshold. The stack includes Mimi, a neural audio codec that compresses 24kHz speech to a 1.1kbps stream at 12.5Hz with 80ms frame latency and outperforms competing streaming codecs. Two fine-tuned voice variants ship: Moshiko (male synthetic voice) and Moshika (female). Weights are available in PyTorch (bf16, int8), MLX (int4/int8/bf16 for on-device Mac and iPhone inference), and a Rust/Candle backend intended for production deployments. Typical workflows are running the browser UI with built-in echo cancellation for demos and research, embedding the Rust server behind a WebSocket for latency-sensitive voice apps, or using MLX to prototype on-device voice interactions on Apple Silicon. Because everything is open weights under permissive licenses, Moshi is best suited to researchers exploring duplex speech modeling, engineers building low-latency voice interfaces without sending audio to a third-party API, and teams that need to fine-tune or study the internals of a full speech dialogue system.
Moshi is the most interesting open release in real-time voice AI — genuinely full-duplex, sub-200ms, and permissively licensed. It won't replace GPT-4o Voice for consumer polish, but if you care about owning your voice stack or studying how duplex speech models actually work, nothing else in the open ecosystem comes close.
— The AI Tool Bible editorial team
Pros
- ✅ Truly full-duplex — handles interruptions, overlap and back-channels rather than rigid turn-taking
- ✅ Sub-200ms practical latency on a single L4 GPU, well below third-party voice APIs
- ✅ Fully open weights (CC-BY 4.0) plus MIT/Apache code — self-host with no per-minute billing
- ✅ Ships with Mimi, a streaming neural audio codec that beats SpeechTokenizer and SemantiCodec
- ✅ Multiple inference backends: PyTorch for research, Rust/Candle for production, MLX for on-device Mac/iPhone
- ✅ Inner-monologue text prediction gives you a transcript alongside the audio stream for free
Cons
- ⚠️ English-only voices at launch — no multilingual support out of the box
- ⚠️ Knowledge and reasoning quality trail top text LLMs; it's a 7B-class model, not GPT-4o Voice
- ⚠️ Requires a 24GB+ GPU for the reference PyTorch build; on-device is only viable via MLX on Apple Silicon
- ⚠️ No hosted API or SaaS tier — you own the ops, scaling and safety filtering
- ⚠️ Only two fixed synthetic voices (Moshiko/Moshika); no voice cloning or speaker conditioning in the release
Facts
- Open source
- Yes
- Model used
- Moshi (7B-class speech-text foundation model) + Mimi neural audio codec, in-house by Kyutai
- Hacker News
- 12 stories · top 635 points · last Oct 16, 2026 · checked Nov 14, 2026
Where this tool sits
Use cases
Frequently asked
- How much does Kyutai Moshi cost to use?
- Moshi is free and open source. Models are under CC-BY 4.0, code under MIT/Apache 2.0. You pay only for your own compute, requiring a 24GB+ GPU or Apple Silicon for self-hosting.
- What hardware do I need to run Moshi?
- You need a 24GB+ GPU for PyTorch inference or Apple Silicon for MLX. Weights are available in PyTorch, MLX, and Rust/Candle backends for production deployments.
- Is Moshi better than traditional voice assistants?
- Moshi processes audio natively without ASR/LLM/TTS pipelines, enabling interruptions and back-channels. It achieves roughly 200ms end-to-end latency on an L4 GPU, below human turn-taking thresholds.
- Can I run Moshi on my Mac or iPhone?
- Yes. MLX weights support int4/int8/bf16 inference for on-device Mac and iPhone. This allows prototyping voice interactions locally on Apple Silicon without external APIs.
- What languages does Moshi support?
- The documentation does not specify multilingual coverage. It is best for researchers and engineers needing self-hosted, low-latency spoken dialogue with open weights, rather than out-of-the-box multilingual support.