Introduction

Zyphra has released ZONOS3-Mini, an open-weight text-to-speech model that clones a voice from a short clip and runs in real time on a phone, a laptop, or straight in the browser. The San Francisco lab published the weights on Hugging Face under Apache 2.0 on 9 November 2026, with a blog post, browser playground, native runtimes, and day-zero serving on SGLang.

What makes this one worth your time is the mix. Small open TTS models usually give up control or quality to stay fast. ZONOS3-Mini keeps voice cloning, voice design by age, accent and texture, and delivery steering down to single words, all with about 100M parameters active per token.

What Shipped

  • Vendor / model: Zyphra, ZONOS3-Mini, the follow-up to the 8B-parameter ZONOS2
  • Modality: text-to-speech with zero-shot voice cloning and an edit model for changing lines after generation
  • Open vs closed: open weights under Apache 2.0; Zyphra's runtime code on GitHub is MIT
  • Runnable today: in-browser WebGPU playground (no install), Zyphra Cloud hosted inference, SGLang-Omni, and Zyphra's own CUDA, MLX, and C++/ggml servers
  • Concrete facts: 800M total parameters with about 100M active per token (64 experts, top-2 routing), 44.1 kHz output, up to 10 minutes of audio per generation, reference clips up to 60 seconds; trained on 112 NVIDIA H100s in 2.5 days

Why It's Fast

Zyphra rebuilt the stack in three places.

  • A new codec. Zyphra Audio Codec 1 (ZAC-1) runs at 15 frames per second instead of the 86 Hz of the Descript codec ZONOS2 used. Zyphra says that cuts the tokens needed per second of speech from 774 to 150, and that under similar data a model on ZAC-1 gets roughly 10x lower word error rate than one on DAC
  • A sparse backbone. Eleven layers mix Gated DeltaNet and global attention 3:1, with latent mixture-of-experts so most of the 800M weights sit idle on any given token
  • A streaming decoder. A causal iSTFT head turns each frame into 44.1 kHz audio as soon as it is generated, so sound starts before the sentence is finished

On one H100, Zyphra reports a median 80 ms to first audio with eight concurrent requests and a peak of 457 seconds of audio per second. On-device, the blog lists under 570 ms to first audio in the browser on a Pixel 9 and under 150 ms on an M3 MacBook Air. The launch thread on X quoted more conservative figures, under 800 ms and under 450 ms, so test it on your own hardware.

Voice Controls

This is where ZONOS3-Mini stands out for creators building characters.

  • Cloning in two modes: drop in a raw reference clip for the most natural result, or use a single speaker embedding for a voice you can then reshape
  • Voice design: steer age, gender presentation, accent (American, British, Australian, and more), and texture such as soft, raspy, bright, or deep
  • Delivery steering: emotion (happy, angry, scared), arousal, pace, projection from intimate to calling across a room, and liveliness
  • Word-level control: whisper or emphasise one span of a sentence while the rest stays neutral
  • Edit model: change a sentence after generation with the same voice and controls

The release ships 27 preset voices across English accents, Hindi, Telugu, French, Italian, German, Spanish, Portuguese, Mandarin, and Korean. Text normalisation covers English, Mandarin, Japanese, Korean, German, Spanish, French, Italian, and Portuguese; other languages pass through as written.

How to Run It

The quickest test is the browser. Open the Zonos Voice Studio, let it download the model once, and clone or design a voice on your own GPU through WebGPU.

To serve it on an NVIDIA GPU with SGLang-Omni:

sgl-omni serve --model-path Zyphra/ZONOS3-Mini --port 8000

curl -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input": "Hello from ZONOS3-Mini!"}' \
  --output output.wav

The endpoint is OpenAI-compatible. Add a reference clip to clone a voice, or turn on streaming to get 44.1 kHz PCM as it is generated. Apple Silicon users can use Zyphra's MLX server, and the C++/ggml build covers CPU, CUDA, Metal, and Vulkan.

How It Compares

Zyphra also released ZTTS2-Eval, a benchmark spanning 35 languages of read speech and 20 languages of in-the-wild recordings, scored across 420 metric, test set, and language combinations. On Zyphra's own index, ZONOS3-Mini ranks first among open models, followed by Fish S2 Pro and MOSS TTS v1.5. The blog lines up side-by-side samples against Fish Audio, Qwen, Cartesia, ElevenLabs, and OmniVoice. It's Zyphra's benchmark, so listen to the samples rather than trusting the ranking alone.

Limits

  • Benchmark and latency claims are Zyphra's own and not yet independently checked
  • The browser build needs WebGPU and a large one-time model download on first visit
  • The emotion controls are targets, not guarantees. Zyphra notes that steering pace or liveliness can also shift loudness and pitch
  • Hosted pricing on Zyphra Cloud wasn't listed in the launch materials
  • Cloning someone's voice still needs their consent. Apache 2.0 covers the weights, not the voices you feed it

Original Source

ZONOS3-Mini announcement on Zyphra's blog

Weights: Zyphra/ZONOS3-Mini on Hugging Face

Code and runtimes: Zyphra/ZONOS3-Mini on GitHub

Try it in your browser: Zonos Voice Studio

Hosted: Zyphra Cloud

Serving guide: SGLang-Omni Zonos3 cookbook

Launch thread:

Conclusion

ZONOS3-Mini is one of the most complete open TTS drops this autumn: Apache 2.0 weights, cloning, a real set of voice and delivery controls, and runtimes from a server GPU down to a browser tab. If you voice game characters, narrate, or build a local assistant, start with the browser playground and try a whispered word or an older, raspier take on your own voice. Then decide whether it's worth a spot in your pipeline next to the closed APIs.

—Anabel ♡