Zyphra drops ZONOS2, an open-source 8B MoE voice cloning model

Zyphra just open-sourced ZONOS2, a text-to-speech model built specifically for low-latency, naturalistic voice cloning. It uses a Mixture of Experts (MoE) architecture with 8B total parameters, but only activates 900M parameters during inference.

The model was trained on more than 6 million hours of diverse multilingual speech. It pumps out high-fidelity 44.1 kHz audio, with Zyphra claiming quality and expressiveness that rivals or beats top commercial TTS platforms.

What it does

ZONOS2 is engineered to clone voices quickly without sacrificing natural inflection.

Language support is split into tiers. Tier 1 covers English, Mandarin Chinese, and Japanese right out of the box. Tier 2 adds 10 languages including French, Spanish, German, and Korean, while Tier 3 spans dozens more, from Polish to Vietnamese.

Under the hood, the pipeline normalizes UTF-8 bytes with nemo TN, extracts speaker identity using ECAPA-TDNN embeddings, and generates discrete audio tokens via DAC tokens using its MoE backbone.

How to run it

Getting ZONOS2 running locally takes real hardware. It is Linux-only on x86_64 and strictly requires an NVIDIA GPU with matching CUDA drivers.

Setup is handled through uv. Zyphra provides an inference server built on Mini-SGLang. Once fired up, it serves a Web UI and API on port 1919 by default.

You can point it at a directory of voice samples in standard formats like WAV or MP3, and the server automatically scans and extracts embeddings for instant cloning. There is also a clean Python API via TTSLLM for batch jobs, or you can try a hosted demo on Zyphra Cloud.

The entire release is available under the Apache 2.0 license, with code on GitHub and weights hosted on Hugging Face.

Why it matters

High-end voice cloning is typically locked behind proprietary APIs with steep per-character pricing. ZONOS2 changes that dynamic by bringing an 8B MoE backbone to local machines under a permissive Apache 2.0 license.

The 900M active parameter footprint is the clever part: you get the expressive headroom of an 8B model with the inference speed of a sub-1B lightweight model. If you are building local voice agents, interactive games, or multilingual dubbing pipelines, this is worth testing immediately, provided you have the Linux GPU box to host it.