Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

ZONOS3-Mini

Zyphra's small open text-to-speech model: 43 languages, voice cloning from a short clip, and fast enough to run in a browser or on a phone.

Reviewed by ReviewRamble editors · Updated

Visit ZONOS3-Mini Open source

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

4.3/5

ZONOS3-Mini puts multilingual voice cloning in a model small enough for a browser tab, under a licence that allows commercial use, which few open speech models offer together.

What is ZONOS3-Mini?

ZONOS3-Mini is an open-weights text-to-speech model with 800 million parameters, of which about 100 million are active per token thanks to a mixture-of-experts design. It speaks 43 languages, clones a voice from a reference clip of up to 60 seconds, lets you design a voice by age, accent and texture, and steers emotion and pace.

It runs in many places: in the browser through WebGPU, on Apple Silicon through MLX, on CPUs and GPUs through a C++ ggml build, on NVIDIA GPUs behind an OpenAI-compatible speech endpoint with SGLang-Omni, and on Zyphra Cloud. It outputs 44.1 kHz audio. Weights are under Apache 2.0.

Is ZONOS3-Mini worth it?

Open text-to-speech has tended to force a choice between quality and size. ZONOS3-Mini narrows that gap: 800 million parameters in total but only about 100 million active per token, so it runs in real time on laptops, phones and in a browser through WebGPU, while covering 43 languages and cloning a voice from a short clip.

Control is better than most open models offer. Beyond cloning, you can design a voice by age, accent and texture and steer emotion and pace down to individual words, which suits games, narration and voice assistants. Output is 44.1 kHz and a single generation can run to 10 minutes. Zyphra reports a median 80 ms to first audio at eight concurrent requests on one H100, and peak throughput of 457 seconds of audio per second.

The deployment story is unusually complete: an OpenAI-compatible endpoint through SGLang-Omni, native servers for CUDA and MLX, a ggml build for CPU, Metal and Vulkan, and 27 default voices. The caveats are that quality and speed figures come from Zyphra's own benchmark, text normalisation covers only nine languages fully, and voice cloning raises the usual consent questions that the model itself does nothing to enforce.

Where it earns its place

  • An indie game studio

    Voice dozens of characters in several languages without studio sessions

    Design voices by age and accent, steer emotion per line, and render on local GPUs; Apache 2.0 allows shipping the audio in a paid game.

  • A developer building a reading app

    Read articles aloud in the reader's language, offline

    Run the WebGPU build in the browser so text never leaves the device; the browser caches the model after the first visit.

Before you commit

Setup
Try it in Zyphra's browser playground, or run sgl-omni serve --model-path Zyphra/ZONOS3-Mini and post to /v1/audio/speech.
Integrations
SGLang-Omni (OpenAI-compatible), Zyphra's CUDA, MLX and C++ ggml servers, WebGPU in the browser, Zyphra Cloud.
Data and privacy
Runs locally or in the browser with no upload; the hosted version is on Zyphra Cloud.
Cost at scale
Free weights; one H100 handles 457 seconds of audio per second at peak, per Zyphra.
Limits
Up to 10 minutes per generation and 60-second reference clips; benchmarks are Zyphra's own.

What’s good

  • 43 languages with voice cloning from up to 60 seconds of audio
  • About 100M active parameters; runs in a browser via WebGPU
  • Apache 2.0, commercial use allowed
  • OpenAI-compatible speech endpoint through SGLang-Omni

What’s not

  • Speed and quality figures are Zyphra's own
  • Full text normalisation in only nine languages
  • No built-in safeguard against cloning voices without consent

Best for

Developers who need multilingual speech or voice cloning on their own hardware, in apps, games or the browser.

ZONOS3-Mini alternatives

ZONOS3-Mini: common questions

Is ZONOS3-Mini worth it?
ZONOS3-Mini puts multilingual voice cloning in a model small enough for a browser tab, under a licence that allows commercial use, which few open speech models offer together. We rate it 4.3 out of 5.
How much does ZONOS3-Mini cost?
Free open weights under Apache 2.0, including commercial use; a hosted version is on Zyphra Cloud, with prices on Zyphra's site.
What are the downsides of ZONOS3-Mini?
Speed and quality figures are Zyphra's own; full text normalisation in only nine languages; no built-in safeguard against cloning voices without consent.
Who is ZONOS3-Mini best for?
Developers who need multilingual speech or voice cloning on their own hardware, in apps, games or the browser.
How much does ZONOS3-Mini cost for a team or in production?
Free weights; one H100 handles 457 seconds of audio per second at peak, per Zyphra.
Is ZONOS3-Mini safe for company data?
Runs locally or in the browser with no upload; the hosted version is on Zyphra Cloud.
What are the best ZONOS3-Mini alternatives?
The closest alternatives we have reviewed are Zonos v2 (ZONOS2).

Answers checked Nov 15, 2026.

Make ZONOS3-Mini? Show this rating on your siteZONOS3-Mini rated 4.3 out of 5 on ReviewRamble