Moshi
The original open full-duplex speech-to-speech model: it listens and talks at the same time with about 200 ms latency. Great for research and natural-sounding chit-chat demos; too small and knowledge-light for most business agents.
Overview
Best for: Research, latency benchmarks, and natural-feeling chit-chat demos on a single GPU or Mac.
At a glance
200 ms on an L4 GPU (160 ms theoretical). 24 GB VRAM for PyTorch; MLX quantized builds run on Apple Silicon. Attribution required.
24 kHz via the Mimi codec (12.5 Hz frames, 80 ms)
24 kHz Mimi-decoded speech
English (not stated on the repo page; the released models were trained for English).
Two fixed voices: Moshiko (male) and Moshika (female).
Vendor claim: 160 ms theoretical, as low as 200 ms in practice on an L4 GPU.
Self-hosted anywhere.
Self-hosted; compliance is your responsibility.
PyTorch: NVIDIA GPU with about 24 GB VRAM. MLX q4/q8 tested on MacBook Pro M3. Rust backend needs CUDA or Metal.
Weights CC-BY-4.0; Python code MIT; Rust backend Apache-2.0.
Features
- full duplex (overlapping speech, backchannels)
- barge-in
- streaming Mimi codec
- PyTorch, MLX (Apple Silicon) and Rust/Candle backends
Pricing
No public price list.
No licence fee; you pay for GPU time.
n/a (self-hosted)
Free tier: Open weights
Source: github.com
Setup
- Linux with an NVIDIA GPU (24 GB for PyTorch) or a Mac with Apple Silicon (MLX).
- pip install moshi (or moshi_mlx on Mac).
- Start the server and open the web UI it serves.
Endpoint
Local: https://localhost:8998 (web UI and WebSocket served by moshi.server)Authentication
None by default; put it behind your own auth/proxy.
Quick start python
# NVIDIA GPU, PyTorch (about 24 GB VRAM)
pip install -U moshi
python -m moshi.server --hf-repo kyutai/moshika-pytorch-bf16
# Apple Silicon (MLX, 4-bit)
pip install -U moshi_mlx
python -m moshi_mlx.local -q 4 --hf-repo kyutai/moshika-mlx-q4
# Rust/Candle server (CUDA)
# cargo run --features cuda --bin moshi-backend -r -- --config moshi-backend/config.json standaloneWritten from the current docs. Check the vendor's SDK version before you ship.
Warnings
Not an agent brain
The 7B backbone has thin world knowledge and no function calling; it is a conversational demo model, not a support agent.
English only, two voices
No voice selection beyond the two fine-tunes and no multilingual support in the main release.
24 GB VRAM for PyTorch
PyTorch quantization is limited (q8 marked experimental); smaller GPUs need the MLX or Candle builds.
Attribution required
CC-BY-4.0 weights allow commercial use but require attribution to Kyutai.
No auth or multi-tenant serving
The bundled server is single-user oriented; batching and auth are your job.
Research cadence
New checkpoints (RAG, RL 'seamless') appear on Hugging Face without stable product docs.
Plus 3 warnings that apply to all open models APIs. See category warnings.
Limits
- Small 7B text backbone: limited knowledge and reasoning
- No built-in tool/function calling
- Context length not stated on the repo
Models and products
| Name | Status |
|---|---|
| kyutai/moshiko-pytorch-bf16 (male voice), kyutai/moshika-pytorch-bf16 (female voice) | GA |
| kyutai/moshika-rag-pytorch-bf16, kyutai/moshika-rl-seamless | Preview |
Docs and sources
Docs
Sources used
- github.com/kyutai-labs/moshi
- huggingface.co/api/models/kyutai/moshiko-pytorch-bf16
- huggingface.co/api/models?author=kyutai
Exact MLX repo name in the snippet; context limit; details of the 2026 RAG/RL checkpoints.