@Gorden_Sun: ZONOS2: Open-source MoE TTS model. 8B total parameters, 0.9B activated parameters. Supports multilingual, voice cloning, Chinese, and Chinese results are good. Model:

ZONOS2: Open-source MoE TTS model with 8B total parameters and 0.9B activated parameters. Supports multilingual, voice cloning, and Chinese, delivering good results in Chinese. Model: https://t.co/ORL0UATU92

— # Zyphra/ZONOS2 · Hugging Face Source: https://huggingface.co/Zyphra/ZONOS2 ZONOS2 title card Discord (https://discord.gg/gTW9JwST8q) — ZONOS2 is our latest text-to-speech model trained on more than 6 million hours of varied multilingual speech, delivering expressiveness and quality on par with—or even surpassing—top TTS providers at low latency with MoE. ZONOS2 excels at high-fidelity and naturalistic voice cloning. During inference we use nemo TN normalized UTF-8 bytes and an ECAPA-TDNN embedding to generate DAC tokens with our MoE backbone. An inference overview can be seen below. ZONOS2 title card Language support is as follows. TierLanguagesTier 1English, Mandarin Chinese, JapaneseTier 2Korean, Russian, Italian, Portuguese, French, Spanish, Vietnamese, German, Hebrew, DutchTier 3Swedish, Hindi, Tamil, Telugu, Thai, Norwegian, Bengali, Tagalog, Arabic, Danish, Indonesian, Polish, Ukrainian, Romanian, Finnish, Hungarian, Lithuanian, Estonian, Slovak, Croatian, Latvian For local inference we provide a high-performance TTS inference server built onMini-SGLang (https://github.com/sgl-project/mini-sglang). For more details and speech samples, check out ourblog (https://www.zyphra.com/our-work/zonos2). We also have a hosted version available atcloud.zyphra.com/audio-playground (https://cloud.zyphra.com/audio-playground). — ## https://huggingface.co/Zyphra/ZONOS2#quick-startQuick Start > Platform Support: Linux only (x86_64). Requires NVIDIA GPU with CUDA toolkit matching your driver version (nvidia\-smito check). ### https://huggingface.co/Zyphra/ZONOS2#1-installation1. Installation Requiresuv (https://docs.astral.sh/uv/getting-started/installation/). git clone https://github.com/Zyphra/ZONOS2.git cd ZONOS2 uv sync ### https://huggingface.co/Zyphra/ZONOS2#2-launch-the-tts-server2. Launch the TTS Server uv run python -m minisgl --model-path Zyphra/ZONOS2 --tts-default-voices-dir ./default_voices/ uv runalways uses the project environment, so no venv activation is needed. The server starts onhttp://localhost:1919by default. TTS mode is auto-detected for zonos2 models.\-\-tts\-default\-voices\-dir pre-populates the web UI with voice-clone speakers from disk; the folder is scanned recursively for speaker audio (\.wav,\.mp3,\.flac,\.m4a,\.ogg,\.opus,\.aac,\.webm) and saved embeddings (\.npy,\.npz). The newest voice is selected automatically on startup. ### https://huggingface.co/Zyphra/ZONOS2#3-generate-speech3. Generate Speech curl: curl -X POST http://localhost:1919/tts/generate \ -H "Content-Type: application/json" \ -d '{"text": "Hello world", "stream": true}' \ --output output.pcm # Convert to WAV ffmpeg -f f32le -ar 44100 -ac 1 -i output.pcm output.wav **Web UI:**Openhttp://localhost:1919/in your browser. ## https://huggingface.co/Zyphra/ZONOS2#python-api-offline-inferencePython API (offline inference) You can also run the engine directly in a Python script, without starting a server, viaTTSLLM: from minisgl.message import TTSSamplingParams from minisgl.tts import TTSLLM tts = TTSLLM(model_path="Zyphra/ZONOS2") results = tts.generate( ["Hello from the offline Python API.", "Batched prompts work too."], TTSSamplingParams(seed=42), ) for i, result in enumerate(results): print(f"frames={len(result['audio_tokens'])}, eos_frame={result['eos_frame']}") tts.save_audio(result["audio"], f"output_{i}.wav") ## https://huggingface.co/Zyphra/ZONOS2#citationCitation If you find this model useful in an academic context please cite as: @misc{zyphra2025zonos, title = {Zonos V2 Technical Report}, author = {Gabriel Clark, Sofian Mejjoute, Mohamed Osman, George Close, Beren Millidge}, year = {2026}, }