What is ZONOS2?
Zyphra's open-source text-to-speech model with zero-shot voice cloning for real-time apps in many languages.
Who is ZONOS2 best for?
ZONOS2 is best for developers building live, multilingual voice apps who want to self-host a TTS model that clones voices.
Faithful voice clones
A new speaker-embedding model captures more vocal detail, and Zyphra reports its strongest scores on speaker similarity and prosody.
Real-time voice apps
Its mixture-of-experts design runs only 900M of its 8B parameters per step, which keeps generation fast enough for live use.
Multilingual and mixed-language speech
It reads text as raw bytes instead of language-specific phonemes, so it handles Chinese, Korean and Japanese better and can switch languages mid-sentence.
Self-hosting under a permissive license
Apache 2.0 weights on Hugging Face and SGLang-based inference code on GitHub let teams run it on their own servers.
How much does ZONOS2 cost?
ZONOS2 is free and open source: Zyphra publishes the weights under Apache 2.0, and it also hosts the model on Zyphra Cloud, which shows no public API prices.
- 8B weights on Hugging Face
- Inference code on GitHub under MIT
- You run it on your own hardware
Zyphra also serves ZONOS2 on Zyphra Cloud, through a playground and an API that need an account, but it publishes no prices for that hosted service. Full pricing on zyphra.com. We checked these prices on .
How do readers rate ZONOS2?
No one has rated ZONOS2 yet. Be the first.
Sign in to rate. One rating per account.
Is ZONOS2 open source?
Yes. ZONOS2's code is public under the Apache 2.0 licence, so you can run it on your own server.
On GitHub
331stars
- Last commit
- 28 Sep
- Repository age
- 6 months
- Licence
- Apache 2.0
- Self-hosted
- Yes
- Repository
- Zyphra/ZONOS2
What are the best ZONOS2 alternatives?
ZONOS2 is an open source alternative to Realtime TTS-2, MisoTTS, Fish Audio S2 Pro and Sonic 3.6. Chatterbox Turbo and Qwen3-TTS do the same job and are open source too.
Inworld's low-latency text-to-speech for live conversation in 100+ languages.
Miso Labs' open-weights 8B text-to-speech model for building expressive English voice agents.
Fish Audio's open-weights text-to-speech model with inline emotion tags, multi-speaker dialogue and 80+ languages.
Cartesia's real-time text-to-speech in 44 languages with voice cloning.
Resemble's open-source low-latency English TTS with voice cloning.
Alibaba Qwen's open TTS models for voice cloning, voice design and preset voices in 10 languages.
Questions about ZONOS2
What is the difference between ZONOS2's stable and expressive modes?
Stable mode aims for clean, studio-like speech and smooths over background noise or distortion in a poor voice sample. Expressive mode copies the reference voice as faithfully as possible, including its quirks, for the most natural-sounding clone.
Who is ZONOS2 for?
ZONOS2 is best for developers building live, multilingual voice apps who want to self-host a TTS model that clones voices.
Is ZONOS2 free?
Yes. ZONOS2 has a free Open weights plan: 8B weights on Hugging Face.
Is ZONOS2 open source?
Yes. ZONOS2's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best ZONOS2 alternatives?
The closest alternatives on Recaply are Realtime TTS-2, MisoTTS, Fish Audio S2 Pro, Sonic 3.6, Chatterbox Turbo and Qwen3-TTS.
Where can I try ZONOS2?
On its official website, zyphra.com.







